R.G. Coyle - System Dynamics Modelling

System Dynamics Modelling
JOIN US ON THE INTERNET VIA WWW, GOPHER, FTP OR EMAIL:

WWW: http://www.thomson.com
GOPHER: gopher.thomson.com
Aservice of ICDP
FTP: ftp .thomson.com
EMAIL: [email protected]
System Dynamics Modelling
A practical approach
R.G. Coyle
Cranfield University, UK
Springer-Science+Business Media, B.Y.

First edition 1996
© 1996 R.G. Coyle
Originally published by Chapman and Hall in 1996.
Soflcover reprint of the hardcover Ist edition 1996
The right of Robert Ge offrey Coyle to be identified as the author of this
work has been asserted by him in accordance with the Copyright and
Patent s Act , 1988.
Typeset in 10/12 pt Times by Best-set Typesett er Ltd., Hon g Kong
ISBN 978-0-412-61710-2 ISBN 978-1-4899-2935-8 (eBook)
DOI 10.1007/978-1-4899-2935-8
Apart from any fair dealing for the purp oses of research or private stud y,
or criticism or review, as permitted under the UK Cop yright Designs and
Patents Act. 1988, this publication may not be reproduced, stored, or
tran smitted, in any form or by any means , without the prior perm ission in
writing of the publishers, or in the case of reprographic reproduction only
in accordance with the terms of the licences issued by the Copyright
Licensing Agenc y in the UK, or in accordance with the terms of licences
issued by the appropriate Reproduction Rights Organization outside the
UK. Enquiries concern ing reproduction out side the term s stated here
should be sent to the publishers at the London address print ed on this
page.
The pub lisher makes no represe ntation, express or implied, with regard
to the accuracy of the information cont ained in this book and cannot
accept any legal respon sibility or liability for any errors or omissions that
may be made.
A catalogue record for this book is availabl e from the British Library
Library of Congress Catalog Card Numb er : 95-71233
~ Printed on perm anent acid-free text pap er, manufactured in accord -

ance with ANSIINISO Z39.48-1992 and ANSI-NISO Z39.48-1984
(Permanence of Paper)
For Julie,
who makes the sun shine,
and for
Rhian and Patricia,
who do the same for my sons.
Contents
Foreword xi
Preface Xlll
1 Introduction to system dynamics 1

System dynamics and management science 1
The paradigm of system dynamics 3
Positive and negative feedback 7
Definitions of system dynamics 9
A structured approach to system dynamics analysis 10
The process of system dynamics 13
System dynamics and top management 15
Summary 16
References 17
2 Influence diagrams 18
Introduction 18
An influence diagram for a central heating system 18
The conventions for influence diagrams 20
The central heating system revisited 22
A production management problem 24
The list extension method 31
The entity/state/transition method 33
The common modules method 36
Which is the right method to use? 39
Some general guidelines for drawing influence diagrams 40
Diagrams at a more aggregated level 40
The cone of influence diagrams 43
Limitations of influence diagrams 45
Level/rate diagrams 45
What is a 'good' diagram? 47
Summary 47
3 Influence diagram case studies 48

Introduction 48
The purpose of a model 48
Vlll Contents
The consulting firm's problem 49

The pharmaceutical company's problem 57
A predator/prey system 62
A simple combat model 67
The decline of the Mayan civilization 73
Extending the case studies 74
Summary 81
References 83
4 Introduction to simulation 84
Fundamental ideas 84
The fundamental equations of system dynamics 86
Time shift and relabelling 89
The simulation process 90
A simulation example 91
Tests of confidence in a model 96
Representing delays in models 98
Smoothing of information 102
Policy experiments with the food model 106
Selecting a value for DT 109
The significance of DT 109
Summary 109
5 System dynamics modelling techniques 111

Introduction 111
TIME 112
A time-dependent process - the RAMP function 113
More time-dependent processes - the sine and cosine functions 114
Sudden changes - the PULSE function 115
Intermittent information - the SAMPLE function 116
Random effects - the NOISE and NORMD functions 117
Logical operations - the CLIP function 119
Limiting functions - MAX and MIN 121
Non-negativity constraints 123
Choosing a suitable delay function 124
Non-linearities - the TA BLE function 130
Model mapping and documentation 136
The complex input pattern 137
Is that function really necessary? 138
Summary 138
6 Case studies in modelling 140

Introduction 140
The domestic manufacturing company (DMC) 140
The management consultant's problem 155
Contents IX
The simple combat model 169

The predator/prey model 180
The rules for time postscripts for variables 192
Summary 194
7 Policy experiments with system dynamics models 195

Introduction 195
Policy in system dynamics models 195
Experimentation versus design 196
Business process re-engineering 197
The domestic manufacturing company 197
Pressure points in the combat mode l 222
Summary 231
8 Optimization in system dynamics 236

Introduction 236
The system response and the parameter plane 237
Hill-climbing optimization 239
Overcoming the limitations of heuristic algorithms 241
The performance of a simple model 242
Formulating an objective function 243
The significance of the objective function 244
Optimization experiments 244
Simulation by repeated optimization 247
Summary 247
9 Optimization in practice 249

Introduction 249
The importance of experimentation before optimization 249
The DMC problem 250
Assessing the value of optimization 271
Constrained optimization 276
Optimization of the combat mode l 277
A matter of priorities 286
Maximum performance or minimum cost? 287
More complex objective functions 287
Factors in choosing an objective function 292
Non-linear optimization 294
Fitting models to data 295
Summary 296
10 Advanced modelling 297
11 The applications of system dynamics 306

x Contents
Introduction 306
Some system dynamics history 306
The New Zealand forestry industry 309
Developing Argentina's electricity capacity 315
Comparison of the New Zealand and Argentina cases 322
Qualitative system dynamics 323
An air traffic control system 323
The drugs 'problem' 326
Business effectiveness - the case of Universal Industries 330
A study of police force measures of performance 331
The collapse of a company 338
System dynamics applied to defence analysis 340
Summary 343
References 343
12 Summary and conclusion 346

Introduction 346
The essence of system dynamics 346
In what circumstances? 348
System dynamics in practice 350
Assessing the quality of a system dynamics model 353
Teaching system dynamics 355
Further reading 358
Finally, on a personal note 359
References 360
Appendix A: System dynamics software packages 361

Introduction 361
Historical pedigree 361
Supplier summaries of the packages 363
Factors in selecting a package 366
Summary 371
Appendix B: System dynamics problems and solutions 374

Introduction 374
Problems 374
Solutions 381
Appendix C: COSMIC supporting documentation and software 404
Appendix D: Using the disk 407

Foreword
More and more people are realizing that social and business systems are too
complex to be understood by intuition , compromise, and superficial debate.
But, the traditional social and managerial sciences are providing little help
in designing policies for better behaviour of large dynamic systems. The
rapidly growing field of system dynamics is increasingly seen as the best
hope for dealing with multiple-feedback-loop, nonlinear systems that ex-
tend across many different intellectual disciplines.
Conferences on the application of system dynamics to business are be-
coming larger and more frequent. Many system dynamics models deal with
interactions between people and the environment. Use of system dy-
namics is under way in economics, government and the management of
universities.
The field needs an expanding literature in each of the many areas where
people are concerned with how things change through time. In this book,
Professor Geoff Coyle has written a welcome addition to system dynamics.
He brings to this book an extensive background as a leader in both the
academic and operational sides of system dynamics. At the University of
Bradford, Coyle founded one of the early academic programs in system
dynamics. The extensive practical aspects of his career are reflected here in
experiences with systems in business and government. His professional
background yields insights regarding both systems and the political and
psychological aspects of working with clients.
Beyond modelling of specific systems, Coyle also is explicit about his
modelling philosophy and raises issues that deserve increased attention and
debate. Some of his modelling practices are of great importance but are
often overlooked, some are controversial and need more discussion, and
some represent unsettled territory. His stress on consistency of units in
equations is a welcome emphasis on a fundamental step in good modelling
that is too often overlooked. Coyle presents an approach to computer-
optimization of models, which is an area of intriguing promise but where
the strengths and limitations are not yet clear.
Geoff Coyle has given us a well-written and useful contribution to the
practical side of system dynamics. He sets an example and opens questions
that can shape future work in the field.
Jay W. Forrester
June 1995
Preface
I have spent more than 25 years using system dynamics for modelling
complex problems, mainly in management and defence, and this book is the
result. System dynamics (SD) has proved to be particularly powerful for
analysing why social, economic, ecological and other managed systems do
not always behave as we wish them to. It sheds light on why management
actions can have negligible or destructive effects, rather than the beneficial
ones which were hoped for, and why systems sometimes behave in ways
which are contrary to intuition.
The system dynamics method has also proved to be particularly good at
supporting a strategic point of view, in the sense of matching very closely the
concerns of top-level decision-makers. By supporting the modelling of the
forces underlying a system's evolution into the future, it allows the analyst
to offer advice on how evolution can be directed so that the system can be
made robust, in that it is enabled to defend itself against setbacks and to
exploit opportunities.
Much of the strength of SD comes from its ability to be used in two
related, but different, ways. On the one hand, it can be used qualitatively to
portray the workings of a system as an aid to thinking and understanding.
On the other, the diagram can be turned into a simulation model for
quantitative simulation and optimization to support policy design.
The purpose of the book is to introduce and explain these powerful ideas,
to illustrate them with some cases and, above all, to teach how these models
are built and used. It is, in short, a model-builder's textbook and the aim is
to bring the reader to the point at which he or she can build, analyse and
optimize system dynamics models of serious problems. As such, it should be
useful on graduate and undergraduate courses in the management sciences,
economics and other disciplines in which the student learns to build and
analyse models of problems involving human intervention in social and
economic systems. It should also be useful to academics in other disciplines
and to practitioners in industry, business and government who wish to
acquire a new skill.
I have long since lost count of how many SD models I have worked with,
but it must be hundreds. They range from simple teaching models through
the work of numerous research students at three universities and into my
own practical work as a consultant to numerous business firms and govern-
ments and in NATO. The experience has convinced me that mistakes are
XIV Preface
very easy to make when one is modelling a complex problem and that it is
easy to overlook errors. The book therefore places much emphasis on
techniques for avoiding errors. The treatment is developed through simple
models to illustrate ideas and more complicated case studies to develop
understanding and skill. Throughout the book important points are printed
in bold type when they are first introduced and italic type when it is
necessary to re-emphasize them .
My ideas about, and understanding of, system dynamics have developed
considerably during the 15 years since I wrote Management System Dynam-
ics. This is, therefore, a completely new work. As with all books , it has
benefited from the comments of colleagues and students too numerous to
mention. I am, however, particularly grateful to Professor Jay Forrester,
who introduced me to SD many years ago. Professor Eric Wolstenholme
was a stimulating colleague with whom I worked happily for a long time.
Many of my colleagues and students at Shrivenham have made helpful
comments on earlier drafts of the book. Mr Ken McNaught and Mr Don
Valler have been meticulous in reading drafts and finding typing errors and
infelicities of explanation. Those that remain are my responsibility .
My thanks are also due to Arlene Curtis and Stephanie Muir who put up
with my occasional distractedness, guarded my door and telephone when I
was trying to work and did endless photocopying in the midst of running a
very busy departmental office.
The people who, above all, have supported the years of work which led to
this book are my wife and sons. When my children were young they, and my
wife, endured many hours of 'Daddy is working'. They, and above all my
wife Julie, have never failed to encourage me when the work wasn't coming
out . My elder son, Jonathan, makes his living from SD and has proved to be
extremely good at it.
RGC
Shrivenham
March 1995
CHAPTER 1
Introduction to system
dynamics
SYSTEM DYNAMICS AND MANAGEMENT SCIENCE
This chapter introduces system dynamics and gives the reader a preview of
what is to come. We start, however, with a history lesson, as it is usually
helpful to be able to put topics into a wider context. In this case, we want to
relate the system dynamics modelling methodology to other approaches to
the study and analysis of managed systems so that we can see how system
dynamics complements and competes with other methods.
During the 1940s, formal analysis, often involving mathematical and
statistical techniques, had been applied to the problems of fighting a war
and, subsequently, to the running of industries and business firms. These
methods were eventually formalized into the disciplines of operational
research and management science, on which there are any number of good
textbooks. For example, the mathematics of linear programming (LP) was
evolved and applied to such problems as how to plan production in an oil
refinery so that given quantities of a range of end products could be manu-
factured at minimum cost by a very complicated process from crude oil
inputs of varying chemical composition and price. This is a problem of
considerable importance to oil companies, and LP, and its later variants, are
very useful for solving it. Similarly, techniques of decision theory were
developed for tackling problems of choice when there are a range of
uncertain outcomes and possibly several conflicting criteria. There are
several more management science methodologies for different categories of
problem.
These methods have proved to be excellent and powerful for dealing
with certain classes of problems: in essence those in which the ability
of a system to adjust to changing circumstances as time passes is not a
significant aspect of the management problem. However, the behaviour of
a system as time passes and new decisions have to be made is a significant
type of management problem which requires the analyst to tackle the issues
of how a system (a term we shall define in a moment) reacts to dynamic
forces and how those reactions shape its behaviour as it moves into the
future .
2 Introduction to system dynamics
A good example of a dynamic system is the ordinary domestic central

heating system. The home owner sets a desired temperature on the thermo-
stat and the system does the rest, controlling itself as time passes and
circumstances change. If the outside temperature falls, the room eventually
cools down to the point where the heat ing is turned on until the house
is comfortable again. If the weather is warm, the system turns itself off until
the house cools down and heat is required again. A more sophisticated
system might also include air conditioning, which will be applied, automati-
cally, to cool the house to the owner's desired setting. A system which
is vastly more complex in operation, but comparable in purpose, is the
automatic pilot on an airliner , which is capable of accurately controlling
the aircraft to follow a pre-set flight plan over great distances despite
the occurrence of unpredictable forces such as adverse headwinds. In
some cases, the aircraft can be controlled during the final approach to
land without human intervention. There are many other cases of the same
type.
These control systems have common features which we shall emphasize
in bold type. They recognize, or sense, the effect the external environment
has produced on the actual state of the system, such as the temperature of
the room in response to the weather outside. They compare the actual state
with the desired state and they employ 'rules' or policies, specifying what to
do in given circumstances, such as, perhaps, that the room can be 2 degrees
cooler than the required temperature before the heating is turned on and 3
degrees warmer before it is turned off. They nearly always involve delays
before the action has an effect; it takes time to warm a house. They depend
on information feedback; when the house is sensed as having warmed up,
the heating goes off.
lt is intuitively obvious that poorly designed rules would give poor per-
formance from a central heat ing system. If the point at which the heating is
turned on is set too low, the house will not be comfortable. If the off point
is too high, heat will be wasted. If the two points are too close together the
heating system will probably not work very well because of the time lags in
the system. However, this does not happen in practice because the skill of
designing control systems has been well learned.
In the early 1960s it was recognized that this control systems viewpoint
could be very powerful if applied to business firms. To take a very simple
example, if demand for its products starts to rise, how quickly should a firm
recruit more workers to increase production? There will almost certainly be
a delay before the new staff can be recruited and trained, so there is a
danger that the work force will overshoot the level required to meet the
apparent increase in demand, not least because the delay in increasing
production may mean that the demand cannot be met and will go else-
where. Even worse, the new staff will have to be paid even before they are
productive and the resulting drain on cash might mean cutting the market-
ing budget, which may be a very good way of killing off the demand. The
firm might, of course, use sub-contractors, though usually at less profit and
with less control over quality .
Under what circumstances, then, should the firm use different policies to
control its behaviour as time passes and circumstances change ? How can its
policies be designed so that the firm becomes robust against change and is
able to create and exploit opportunities and avoid, or defend itself against ,
setbacks? How can the organization design its information feedback struc-
ture to ensure that effective policies become possible? These are the issues
that system dynamics addresses.
In practice, most problems, whether they are in business, economics , the
environment or whatever, are much more complicated than the workforce
expansion problem discussed above . The reader might care to pause for a
moment and jot down a few more of the influences which might arise in that
problem, just to get a feel for how complex it can become . Because prob-
lems can be so complicated, it is necessary to have a methodology for
dealing with them, and that is what system dynamics provides and what this
book will explain . System dynamics can, therefore, be interpreted as the
branch of management science which deals with the dynamics and control-
lability of managed systems .
The recognition that this type of problem exists in the business world, and
that the concepts of control theory could be applied, is due to Professor Jay
Forrester, of the Massachusetts Institute of Technology. He further realized
that the mathematical techniques of control theory do not apply to man-
aged systems because they tend to be much more complicated than engi-
neering problems. The final stage in Forrester's intellectual tour de force
was, therefore, to define the structure of specialized computer simulation
languages to enable the calculations in system dynamics to be performed
quickly and easily.
Since Forrester's pioneering steps, some 35 years ago, system dynamics
has been applied to an enormous range of problems, some of which will
be reviewed in Chapter 10. For the present, we must get on with learning
about what system dynamics does, how it works, and the techn iques
required.
THE PARADIGM OF SYSTEM DYNAMICS
The paradigm, or basic viewpoint and associated methodologies, of system

dynamics requires that we first define a 'system'. There are many definitions
of this elusive idea, most of which imply that a 'system ' is a collection of
parts which interact in such a way that the whole has properties which are
not evident from the parts themselves . Thus, the behaviour of an angry
human being cannot be explained from the chemical composition of the

body, or, at any rate, not yet. Such definitions sometimes tend towards an
almost mystical stance, and we shall adopt a more pragmatic point of view
and define a system as:
a collection of parts organized for a purpose
This is simple and lays emphasis on the ideas that there are 'parts', such as
labour recruitment and marketing in the earlier example, which are 'organ-
ized', in the sense that they relate to each other in some way, and they
attempt to achieve or contribute to a purpose, such as the survival and
prosperity of the firm. This is all very obvious and the real subtlety lies in
the idea that the system may fail to achieve its purpose. It may fail because
its parts are poorly designed, because they are poorly connected, because it
is knocked off course by an external shock, because the purpose is inher-
ently incapable of being achieved or because its attempts to adjust to
change, its policies, are badly designed, even to the extent of making mat-
ters worse.
In principle, any of these failures could arise in a central heating system,
but they rarely do because such systems are well designed. The fundamental
purpose of system dynamics is to achieve comparable quality of design, and
hence performance, in managed systems.
System dynamics achieves this purpose through adopting the viewpoint
shown in Fig. 1.1 and supporting the viewpoint with a systematic procedure,
which is covered later in this chapter, and a range of techniques which form
the subject matter of the rest of the book.
Information
-~- "
/
State
/'
'" -,
\
~
en Knowledge
(1)
o \
c \
(1)
:::J
0- I
(1)
I»
o
en
c
o IJc
/
g.
:::J
/
/'
,./
Choice . . -
Fig. 1.1 The information/action/consequences paradigm of system dynamics.

Figure 1.1 can be understood by first following the loop . This shows that
the 'state' of the system, such as the size of the workforce, derives from
'choices' made by its managers, such as recruiting more people, making
them redundant or not replacing those who leave. The large D on the link
from Choice to State reminds us that there is usually a substantial delay
between a choice being made and the effect being felt in the state. The
subscript on the D simply makes it clear that the delays in the different
stages may be different. The next step is that the 'state' becomes known to
the controllers of the system, perhaps in the form that people are leaving at
an average rate of 20 per week from natural wastage. There will, however,
be a knowledge delay before the fact that people actually are leaving at the
rate of 20 per week becomes realized and accepted. This knowledge then
leads, after a delay for choice, to new actions, such as to attempt to recruit
30 per week, because it is desirable to increase the workforce smoothly, but
not by too much.
The words round the outer edge emphasize the information/action/
consequences paradigm. Information produces actions which have conse-
quences, generating further information and actions, and so on. This se-
quence of the three components as time passes is dynamic behaviour. The
state will change continually as time goes by and the pattern of behaviour
will depend on how well the information and actions are tuned to each
other and to the way in which consequences arise in the particular system.
The tuning must take account of the delays . The techniques of system
dynamics are the tools by which the tuning is achieved to make the pattern
of behaviour as acceptable as it can be.
The second point about Fig. 1.1 is that the central loop is a loop; that is
that there is a closed chain of cause and effect in which information about
the result of actions is fed back to generate further action. These feedback
loops are a central element both of engineering control theory and of
system dynamics. This connection now allows us to offer a preliminary
definition of system dynamics as:
the application of the attitude of mind of a control engineer to the
improvement of dynamic behaviour in managed systems.
This involves analysing a managed system so as to:
• model the ways in which its information, action and consequences com-
ponents interact to generate dynamic behaviour
• diagnose the causes of faulty behaviour
• tune its feedback loops to get better behaviour.
A control engineer uses sophisticated mathematical tools, supported by
computer simulation, to go about those tasks for an engineering system .
System dynamics makes use of control theory insights, simulation and
optimization to deal with managed systems.
The tuning of a managed system means adapting its policies and the
structure of its loops to achieve better behaviour. The dotted lines for
information and action in Fig. 1.1 imply that these parts of a system are
capable of being changed. As we shall see when we analyse models, loops
can be abolished or created, and it is this ability to redesign parts of the
system which makes system dynamics a viable methodology for managed
systems.
It is tempting to assume that a managed system is always a business firm,
and there have indeed been numerous examples of the study of such sys-
tems. System dynamics has, however, also been applied to the dynamics of
decline in cities, to oscillation in ecological systems, the growth and collapse
of civilizations, the behaviour of national social and economic systems, the
future of the world, the criminal justice system, and to a range of military
problems as well as to many other topifs . All of these have two things in
common.
The first is a wish to make matters better by suggesting how people can
act upon the system, which is why we term them 'managed systems'. What,
for example, would be a good policy for controlling fishing so that food is
obtained but stocks are not depleted? How can the criminal justice system
be made more 'effective'? What type of labour recruitment policy should a
firm adopt? These questions all imply that there must be a set of criteria by
which we shall assess the quality of performance. Evidently, criteria are not
a simple matter, as some aspects may be more important than others and it
may be necessary to trade off one against another. We shall have much
more to say about this in the chapters on optimization.
The idea of trading off between criteria means that we must know what
is supposed to be achieved, and that may depend on the circumstances. A
firm which is experiencing growth in demand may wish to expand as fast
as possible, though without over-taxing itself and getting too big for the
market. If, on the other hand , the same firm faced a declining market it
might well seek to contract in size in an orderly fashion. The required
behaviour in these two cases would be different, but it would be desirable
for the behaviour to be the best that could be achieved in the circumstances.
This is expressed in the important principle of robnstness, which states
that:
the system should always perform as well as the circumstances allow,
regardless of what the circumstances are.
In particular, a managed system should be able to defend itself against,
or recover from, shocks, and it should be able to create and exploit
opportunities.
The second thing that managed systems have in common is the ubiqui-
tous presence of feedback loops in all these systems. There are, however,
two types of feedback loop: goal-seeking, or negative loops and growth-
producing or positive loops.
Positive and negative feedback 7
POSITIVE AND NEGATIVE FEEDBACK

The differences between the two types of loop are quite crucial in under-
standing dynamic behaviour.
Central heating systems, aircraft autopilots and any other system, includ-
ing managed systems, which seeks to achieve a target must embody at least
one negative or goal-seeking feedback loop. The term 'negative feedback'
is the standard one and is used throughout this book, but goal-seeking is
what the loops try to achieve'. The standard form of a negative loop is
shown in Fig. 1.2(a).
The essential idea of negative feedback is that, when there is a difference
between the desired and actual states of the system, actions are generated,
Desired _
State
-.... Discrepancy
.,.. . . . Between Actual .......
and Desired '- System
/
"' Policies
ExtShocemksal _ _....... Actual
State \ I
\
Action
Completed
\
+ ;'
~;t~~.
a Negative Feedback
~~::,;;f_
<,
-,
\
\
I
I
/
/
Rate of ... . /
Change
Change'" - -- Parameter
b Positive Feedback
Fig. 1.2 Negative and positive feedback loop structures.
1 Some authors refer to negative loops as 'balancing ' loops. This is acceptable , though it
implies that the loop will succeed in its balancing act. A badly-designed negative loop can,
however, unbalance a system, so the term goal-seeking is probabl y better. The traditional term
of 'negative feedback ' will be used throughout this book.
under the influence of the system's policies, so as to attempt to eliminate the

difference ; the loop seeks the goal of having no difference between desired
and actual conditions and the quality of the performance of the system as a
whole is assessed by how well that is achieved as time passes.
In a simple case, such as a central heating system, the desired and actual
states are the temperature set on the room thermostat and how warm
one feels. If the two match quite closely as time passes, despite rises and
falls in the outside temperature, the system is performing well. In another
example, a football team has a desired state of aiming to win matches and
an actual state of the matches it does win. Its policies would affect the
choosing of the team for successive games or the purchase of new players .
If the policies are successful, the team will perform well by being close to
achieving its targets throughout the season . An ill-judged policy of changing
the team before it has had a chance to settle down might easily make
matters worse.
In Fig. 1.2(a), the policies are shown in italic, suggesting that they are
fixed 'rules' specifying what is to be done in given circumstances. Later,
we shall meet policies which vary, in the sense that they may be chosen from
a menu of options . However, there is nearly always a delay before the
action which was started (turning on the central heating boiler or starting to
recruit workers) is completed, and it is the completed action which affects
the actual state of the system. That state is, however, affected , perhaps
continually and before actions can take effect, by external shocks such as
changes in the weather or another firm offering better wages. The diagram
shows the influences at work in negative feedback and is, in fact, a simple
version of the influence diagrams which we shall study in the next chapter,
at which point we shall introduce + and - signs on the links to give a much
more precise meaning to terms such as 'actual state ' and 'discrepancy'.
Notice, though, that solid lines represent the consequences parts of the
paradigm and dashed lines represent feedback of information and the
generation of actions.
Ideally, negative feedback should produce a stable reaction to external
shocks. If a discrepancy arises, it should be smoothly eliminated within a
reasonable period of time. In fact, unless the policies are well designed in
relation to the 'physics' of the consequences in the system, the loop may not
succeed, and a discrepancy may oscillate between the workforce being too
large and too small and the oscillations may go on forever. In severe cases,
the oscillations may get progressively larger!
Positive feedback, on the other hand, acts as a growth-generating mecha-
nism. The state of the system, such as one's bank balance , grows continually
larger as interest payments act as the rate of change . The change parameter
is the percentage interest paid by the bank. This is sometimes called a
virtuous circle, as opposed to the vicious circle which arises when the
balance becomes negative and one gets deeper and deeper into debt as
interest is added to the debt. Positive feedback is quite common in managed
systems and may be valuable as an engine of growth. In an engineering

system, however, positive feedback is undesirable and is designed out,
which is one reason why the mathematical techniques of control engineer-
ing are of little help in designing managed systems.
DEFINITIONS OF SYSTEM DYNAMICS
Having discussed some basic ideas, we are now in a position to put forward
some definitions of system dynamics. The earlier remark about the applica-
tion of control engineering to managed systems makes it clear that the aim
is to achieve in managed systems the same standards of performance that
control engineers can attain in very complex systems such as the autopilot
for a Boeing 747. It does not , however, tell us much about how system
dynamics goes about its task.
Forrester, the founder of system dynamics, defined it as (Forrester, 1961):
... the investigation of the information-feedback characteristics of
[managed] systems and the use of models for the design of improved
organizational form and guiding policy.
In a previous book, this author's version was (Coyle, 1979):
A method of analysing problems in which time is an important factor,
and which involve the study of how the system can be defended
against, or made to benefit from, the shocks which fall upon it from the
outside world.
or, alternatively
System dynamics is that branch of control theory which deals with
socio-economic systems, and that branch of Management Science
which deals with problems of controllability.
Wolstenholme (1990) offered:
A rigorous method for qualitative description, exploration and analysis
of complex systems in terms of their processes, information, organiza-
tional boundaries and strategies; which facilitates quantitative simula-
tion modelling and analysis for the design of system structure and
behaviour.
None of these is completely satisfactory. Forrester does not say what type of
models are involved. Neither Forrester's nor Wolstenholme's definitions
refer to time, and mine does not mention information feedback.
Wolstenholme rightly mentions qualitative analysis but does not include
the powerful optimization methods that we shall study later in this book .
Perhaps something on the following lines will suffice to start us on our
study:
System dynamics deals with the time-dependent behaviour of managed

systems with the aim of describing the system and understanding,
through qualitative and quantitative models, how information feed-
back governs its behaviour, and designing robust information feedback
structures and control policies through simulation and optimizat ion.
That is a shade on the long side, but readers should accept it for the
moment. At the end of the book we shall challenge you to define system
dynamics for yourself.
A STRUCTURED APPROACH TO SYSTEM DYNAMICS

ANALYSIS
Having discussed some of the underlying ideas of what system dynamics is

about, we turn to the underlying five-stage approach for its application,
which is shown in Fig. 1.3.
The first stage is to recognize the problem and to find out which people
care about it, and why. It is rare for the right answers to be found at this
stage, and one of the attractive features of system dynamics as a manage-
ment science methodology is that one is often led to re-examine the prob-
lem that one is attempting to solve.
Secondly, and the first stage in system dynamics as such, comes the
description of the system by means of an influence diagram, sometimes
referred to as a 'causal loop diagram '. This is a diagram of the forces at work
in the system which appear to be connected to the phenomena underlying
people 's concerns about it. Influence diagrams are constructed following
well-established techniques which will be explored in the next chapter. The
arrow that leads from Stage 2 to Stage 4 will be explained in a moment.
Having developed an initial diagram, attention moves to Stage 3, 'quali-
tative analysis'. The term simply means looking closely at the influence
diagram in the hope of understanding the problem better. This is, in prac-
tical system dynamics, a most important stage which often leads to signifi-
cant results. Indeed, the problem is sometimes solved at this stage and there
is no need to go on to the others , which is the meaning of the dotted line
between Stages 3 and 4.
In Stage 3, the analyst draws on so-called bright ideas and pet theories.
The former arise from experience with other problems. One may have seen
something like the set of feedback loops for this problem in some other
case, and what was learned then may be applicable here. That, of course,
does not help the inexperienced analyst who may have to rely on looking
for obvious inadequacies in the system. If, for instance, it contains negative
feedback loops which have no clear desired states then it is very unlikely to
be successful in eliminating discrepancies and may well be showing evi-
dence of being out of control; managers talking about never having the right
A structured approach to system dynamics analysis 11
Sta ge 1 Problem Recognition

(who cor es, a nd why )
Stage 2 Problem Understanding and System Description

(Influenc e Diag rams)
Sta ge 3 Qualita tive Analysis

(bright idea s and pe t theo ries) ..,
Stage 4 Sim ulation Modelling

(spe c ia l computer simulation languages)
Model testing
Stage 5 Policy Testingand Design .... ---1
Sensitivity testing
,
5A ExploratoryModelllngand ltt.'nsights
..
Policy Designby Simulation ~ Ideas
(assessment by jud gement)
(objec tive function)
.. Robust
56 Policy Design by Op timisation --~'~POlicieS
Fig. 1.3 The structure of the system dynamics approach,
numb er of people available could be a symptom of that. Similarly, a system

which should be growing ought to have positive loops. If they cannot be
found in the influence diagram then a diagnosis of the probl em may be
emerging.
Pet theories are frequ entl y even more useful. They are the views of
experienced people in the system as to what is wrong with it. The views
themselves may be found to be wrong on deeper analysis, and the reasons
why they are wrong are usually of great interest, but they are almost always
a useful source of knowledge about the probl em and should be searched for
by the analyst.
If qualit ative analysis does not produce enough insight to solve the prob-
lem, work proceeds to Stage 4, the construction of a simulation model. At
this stage, we exploit the important prop erty that , as we shall see in Chap-
ters 2 and 3, the influence diagram can be drawn at different levels of
aggregation. It is usually not even necessary to show every single detail,
because, if the influence diagram has been properly drawn, the simulation
model can be written from it without a separate stage of flow charting. This
derives from the use of special simulation languages for system dynamics
and , in effect, the influence diagram and the simulation model are simply
two versions of the same model; one written in arrows and words, the other
in equations and computer code. This property is of fundamental
importance in system dynamics as it gives rise to some powerful practical
consequences.
The first is speed . The equations can be written in any order the modeller
chooses. In practice, an order is chosen which is easy to understand and
explain, but the modeller does not have to worry about the order in which
calculations must take place; the specialist languages do that. This means
that the model can usually be written 10 or even 100 times faster than the
equivalent model in BASIC or any other high-level programming language .
The time saving means that initial results from a model can often start to
emerge within days or even hours of starting work, which is vitally impor-
tant from the point of view of communicating with, and retaining the
interest of, a client or sponsor.
The second is ease of revision and expansion of a model. It is a truism of
practical modelling that one never understands the problem correctly at
first and the initial attempt at a model will usually be inadequate in some
way. Since equations can be written quickly, a model can easily be revised
as better understanding develops. Similarly, it is easy to extend a model
when the need arises without the existing code ceasing to work , something
which almost invariably happens when a program written in a high-level
language is extended or modified. Indeed, system dynamics modellers usu-
ally exploit this property by building small models to start with and allowing
them to expand in a controlled fashion, rather than seeking to write a large
model from the outset.
Thirdly , the fact that the model exists in the two equivalent forms of a
diagram and a set of equations is a powerful aid to thinking about and
understanding a problem. One uses whichever form is most conducive to
effective thought and communication.
Naturally, this stage also includes the testing and debugging of the model;
most important steps, which are often referred to as 'validation'
of the model. This is a most inappropriate terminology as it implies
that the model somehow becomes absolutely correct and true. It is far
better to regard 'valid' as meaning : well suited to a purpose and soundly
constructed.
To test the soundness of construction, system dynamics modellers make
use of techniques such as dimensional analysis, mass balance checks, ex-
treme value tests, and other approaches which we shall encounter when we
build some models in later chapters. For the moment, the reader need
simply be aware that testing a model to ensure that it is well constructed is
something which is taken very seriously in system dynamics and that there
are several approaches to carrying out the tests.
The process of system dynamics 13
Stage 5 is where results based on quantitative analysis start to emerge.
Initially, use is made of the insights from the bright ideas and pet theories
from qualitative analysis, which is the meaning of the line joining the two
stages.
In the first stage, the emphasis is on policy design and system testing by
simulating potential changes to the system to see what effect they have and
on trying to understand those effects by thinking about the system's loops.
The assessment of whether behaviour has improved , and why, is done by
qualitative judgement and looking at the output. This is, as it were, Stage
SA, in which the emphasis is heavily on exploring the behaviour of the
system rather than predicting precise details. This stage represents explora-
tory modelling of the system's characteristic patterns of behaviour with the
aim of enhancing understanding.
The large arrow going down to the words 'objective function' suggest
that, as one's insights into, and ideas about, the problem develop during this
simulation stage, it should be possible to develop an extra set of equations
which capture the essence of what the system is trying to achieve, or what
one would like it to achieve, and measure how well it succeeds. Objective
functions exist in other branche s of management science, such as linear
programming, and system dynamics uses them to unleash the power of
optimization software to design robust policies which addres s the issues
identified in Stage 1. This phase of system dynamics is applicable whether
the system is closely managed, as in a business firm or a defence problem, or
whether the purpo se is research into how a past civilization might have
avoided collapse.
THE PROCESS OF SYSTEM DYNAMICS
Figure 1.3 showed the technical stages in the system dynamics methodol-
ogy, but it is also useful to look at the same ideas in another way to
emphasize the relationships between stages of work and the results to which
they give rise. This is shown in Fig. 1.4.
The left-hand side shows 4 of the stages of system dynamics and the right-
hand the results produced. Notice that Stage 3, qualitative analysis, leads
from system description to understanding and ideas and that the latter may
lead back to the problem definition stage. Similarly, Stage SA corresponds
to the use of the model to verify ideas generated earlier and to stimulate
new ones. Finally, Stage SB, Optimization, uses the model as its basis, but
relies more heavily on the emerging ideas and insights which have devel-
oped down the right-hand side.
The double-headed arrow between optimization and robust policies is
intended to emphasize that optimization does not automatically produce
the single answer to the problem. Very often the nature of the policies
suggested by the optimizer leads to further understanding of the problem
14 Introdu ction to system dynamics
Stage I Result
I
I
---------T---------
Problem Definition---...I._
SA
4 Simulation Model......f--.=:..:._+----'.~
(egCOSMIC)
I
Verification, ideas
1
58 Policy Design ......I----r--.~ Robustpolicies
(eg COSMOS optimisation)
Fig. 1.4 The process of system dynamics.
and the workings of the system and hence to a richer vein of thinking
leading to more profound policies being discovered.
Technically, as we shall see in a later chapter, the optimizer works by
running the simulation model many times. This can be called optimization
by repeated simulation . However, the double headed relationship by which
optimization can lead to better optimization is similar to a simulation ex-
periment making one think of a better experiment, and should make one
think in terms of simulation by repeated optimization.
Figure 1.4 refers to two of the specialist system dynamics software sup-
port tools, COSMIC and COSMOS: the former for model building and
testing and the latter for optimization. There are several commercial pack-
ages available and it must be emphasized that the most significant aspect of
successful system dynamics is intelligent and careful use of the package to
which one has access, not the use of any particular package. The models in
this book were developed with COSMIC and COSMOS, but users of other
packages should have no difficulty in using the models, with minor amend-
ments. The software packages available for system dynamics, some of which
do not optimize, are reviewed in Appendix A.
System dynamics and top management 15
SYSTEM DYNAMICS AND TOP MANAGEMENT
We have seen several times that system dynamics concentrates on the

policies and dynamic behaviour of the system in question, which are the
strategic concerns of the top managers, whether they be chief executive
officers of businesses, city managers, generals, politicians or whatever.
Figure 1.5 brings together the characteristics of system dynamics, which
we shall study as the book progresses, and these top management
concerns.
The diagram can be read from left to right or vice versa. Thus, the system
viewpoint of system dynamics and the 'whole company' concern of a senior
manager equate to one another. In particular, the transparency of the
influence diagram and the simple simulation technique provide for the top
manager to understand what is going on. This is precisely the point made
earlier about the duality of the influence diagram and the simulation model
as two representations of the same problem.
The last point in the diagram is worth expanding upon. Quite large
system dynamics models typically take only a minute or two to compile and
run on a standard PC, Making the 20 to 30 runs usually required for an
optimization experiment normally takes only another minute or so. Such
fast model running capabilities are ideal for study periods in which the
Top management
SD Characteristic
concerns
System viewpoint Whole company
Feedback analysis Consequences of actions
Concern with Mure

Dynamic modelling Testingof Ideas
Robustnessagainst
I
Optimisation
uncertainty
Transparency of Understanding. input

Innuence Diagram
and control
Simple s1mulotlon
technique
Fast s1mulal1on Study periods
Fig. 1.5 System dynamics and top management.

client team , with some assistance from an analyst, can work with the model
themselves , testing their own ideas and developing their own insights, over
a period of a day or so. Practical experience of these sessions suggests how
enormously valuable they can be, both in drawing out ideas and insights
from the problem proprietors and in giving them confidence that they
understand and support what has been done and that 'their' problem has
not been taken over by the analyst.
SUMMARY
This chapter has been an introductory survey of what system dynamics is

for, how it relates to the rest of management science, its main processes and
how it supports top level decision issues. It should have given the reader an
idea of what to expect in the rest of the book. A number of claims have been
made about system dynamics, which, in essence, are as follows:
• System dynamics is relevant to dynamic problems and therefore comple-
ments other management science approaches, which are more closely
geared to static problems. The implication is not that static analysis is
inferior or inappropriate; it is that dynamic problems are also important.
• System dynamics deals with the broad behaviour of the system and how
it influences its own evolution into the future . These can be seen as the
strategic issues which concern top management in the organization.
• System dynamics, like all good methodologies, has a systematic proce-
dure ; the five-stage method.
• System dynamics models come in two equivalent forms: the influence
diagram and, where appropriate, a simulation model. These two forms
make it possible to communicate easily and effectively with the client or
sponsor of the work.
• The simulation technique of system dynamics makes it particularly easy
and quick to build models.
• Once the model has been built and tested, the analyst can use it to test
alternative policies and to redesign the system so that its policies become
more effective. Powerful techniques of optimization support this process
but do not replace thinking about what is going on in the model.
• System dynamics can be applied to a very wide range of problems.
• Last, but by no means least, as an intellectual pursuit and as a way
of making a living, system dynamics is very challenging and enormous
fun.
So far, all of these are no more than assertions which will be developed in
detail in the ensuing chapters. Readers should, however, be constantly
reminding themselves of them and continually evaluating them as they
proceed.
References 17
REFERENCES
Coyle, R.O. (1979) Management System Dynam ics, John Wiley & Sons, Chichester.
Forrester, J.W. (1961) Industrial Dynamics, MIT Press, Cambridge MA.
Wolstenholme, E.F. (1990) System Enquiry, John Wiley & Sons, Chichester.
CHAPTER 2
Influence diagrams
INTRODUCTION
In Chapter 1, we stated that influence diagrams are an essential tool in

system dynamics; they are not only the foundation on which quantitative
models are built but are also a valuable device in their own right for
describing and understanding systems. This chapter will cover the tech-
niques for constructing diagrams, and we shall first study a simple example
so as to show the detailed conventions used in influence diagrams. Later, we
shall show how diagrams can sometimes be simplified and study the impor-
tant point that it is not only permissible, but also often desirable, to draw
several diagrams at different levels of detail for the same problem.
Influence diagrams are sometimes called 'causal loop diagrams'. There is
little or no difference, but causal loop diagrams are best thought of as
influence diagrams drawn at a very broad level, and not showing the fine
detail which can be included in an influence diagram . We shall discuss this
in more detail later in the chapter when we consider the 'cone' of influence
diagrams.
AN INFLUENCE DIAGRAM FOR A CENTRAL

HEATING SYSTEM
We start by studying the completed diagram for the central heating system
discussed in Chapter 1, the diagram being Fig. 2.1. Initially, we shall not
worry about the solid and dotted lines, the + and - signs or any of the
details, but concentrate on what the diagram shows, which are the influ-
ences at work in the system, the interplay of which is the cause of its
dynamic behaviour.
For this system, the influences are fairly simple, or, at any rate, readers
will think so when they have studied some of the more complex cases in this
book. The householder has set a desired temperature on the thermostat,
but the actual temperature may not be at that value. If it is not , the boiler
is automatically turned on and heat starts to flow into the water pipes and
radiators or into the hot air ducts. After a delay due to the length of the
pipes, heat starts to reach the room, adding to the quantity of heat already
An influence diagram for a central heating system 19
Hr:..,:.::::::r::t:
"
Desired Temperature
on Thermostat
- \
"
+ SyoIem
'PoIIey'
" <, ,/
- - +~~f':~~"6~-
E,
_ <, /
/
~
_ _ Actu~~T=rajure
/ \ \ /
+, /
,/ +.... /
/ ~ ... -... I \
/ Negative', I
1 Feedback ,
I \
Loop 2 I
\
Negative
Feedback
Rate of Supplying
Heat
+t ' / Loop I ~
Difference Between - - \ \ \
. . _. \
Internal and External Quantity of Heat \ \
~~~a\res _/lnRoom~'
/ + PowtHol
\ ~~
_~'--..., ~Rate of Losing
.- ....", Heat
External ... Rate of Arrival
Temperature .\ of Heat In Room
+
\
Temperature DIIIerence to
Rote 01 Loa 01 H_ ConvetSlon Foetol
Fig. 2.1 Influence diagram for central heatingsystem.
there. Depending on the size of the room , the heat quantity is sensed as
temperature by people in the room and by the thermostat and , when the
temperature is close to the desired value, the supply of heat stops . We shall
suppose that this is also an air-conditioning system so, if the room is too
warm, the system starts to subtract heat until the room cools down. Sub-
tracting heat should be thought of as supplying negative heat , so the words
on the diagram do not need to be changed.
This simple system shows the essential trick in developing a good dia-
gram - 'think physics'. In the information/action/consequences paradigm
the consequences are always physical in the sense that something flows in
the system - people are recruited, heat is supplied, goods are manufactured,
casualties are incurred, money is transferred and so on. Even in a psycho-
logical case, unhappiness builds up or decreases. In all cases, therefore,
something flows, and recognizing what flows in the system is the key to a
good model, a good trick being to visualize the problem as a set of pipes
and tanks through which water flows. In fact, flow analysis is the basis of one
of the techniques of diagram construction which we shall examine later. We
shall leave the question of what a 'good' diagram is until later in the chapter,
when we have built some influence diagrams .
At the risk of labouring the point , in this case the heating system cannot
supply temperature; it must supply heat, and the amount of heat which
has accumulated in a room of a certain size is what determines the tempera-
ture, and it is temperature which the managers of this system wish to
achieve.
20 Influence diagrams
Unfortunately, the higher the room temperature relative to the outside

temperature, the greater will be the rate of loss of heat from the room.
Again, if the external temperature is higher than that in the room, heat will
flow in. That is a negative loss rate so, again, the words on the diagram are
appropriate.
THE CONVENTIONS FOR INFLUENCE DIAGRAMS
It is now time to turn to the diagrammatic conventions. First, we shall state

them, and then explain them in relation to Fig. 2.1.
• Solid lines show physical flows , and these are the consequences in the
paradigm.
• Broken lines represent influences which are not physical flows and in-
dicate the information and action parts of the paradigm. The actions may
be deliberate choices by the managers of the system or they may be the
ungovernable reactions of nature to the state of the system.
• The large D on the physical flow represents a significant time delay.
A subscript may be attached to the D to distinguish one delay from
another.
• A box denotes an external driving force over which the system has no
control and to which it must respond. In this case, the External Tempera-
ture is such a driving force, changing from time to time for reasons of its
own. The Desired Temperature on the Thermostat is also a driving force.
It is set by human action but, having been set, the central heating system
simply has to respond to whatever it is.
• A + sign means that when the variable at the tail of the arrow changes,
the variable at the head always changes in the same direction. Thus, the
greater the quantity of heat in the room the higher the temperature will
be. On the other hand, if the quantity falls, so does the temperature.
• A - sign has the opposite effect: if the tail variable changes then the head
variable changes in the opposite direction . Thus, the greater the rate of
losing heat the less heat will remain in the room and, the smaller the rate
of losing heat, the more heat will be left.
• Items in bold italics are constant factors, usually called parameters.
(Underlining can also be used, especially when drawing by hand. The
choice is immaterial , as long as a parameter is identified as such.) The
dotted arrows leading from them emphasize that they are not physical
flows, they are simply influences on the system. One might use a different
pattern of broken line, but that would complicate the drawing of dia-
grams.
• It is not always obvious whether a parameter has a positive or negative
effect on the variable it influences and, for that reason, signs on links
from parameters are optional. The author sometimes does not include
The conventions for influence diagrams 21
them in diagrams, though that is partly a matter of personal style, and
readers should develop their own styles as skill improves with practice .
e Feedback loops have attention drawn to them by a broken circle and a
name for the loop. A loop exists when, starting from a given variable and
following arrows in the direction they lead, it is possible to get back to the
start, without going through any variable more than once.
e The sign, or polarity, of a loop is, theoretically, found by multiplying its
signs together. Multiplying the signs on Loop 1, starting from Quantity of
Heat in Room gives:
+x-x+x+ x+
which, by ordinary mathematics, is negative. A simpler method is to count
the negative signs. If there is an odd number of - signs, the loop is negative,
or goal-seeking; otherwise it is a positive, growth-generating, loop.
These conventions are summarized in Fig. 2.2 with the addition of the
averaging, or smoothing of information. In some systems, a variable, such as
Types of link In an Influence Diagram

-----i~~ Physical Flow 1 Unkshave. ex• ~gN .
Information transmission
---+ Conl rol ocnco
Behaviour of nature
SIgnmay be omllled, bul •

ex. may also be used
Delays In Physical Flow
o denoles delay, subscript

Identffles a particular delay,
sign Is always •
.
Smoothing of Information
"'ARIABLE :..AVERAGE VALUE

.. - - - p- OFVA~A~E
./
./
UMEOVfR
WHICH VARfA~
ISAVfRAGED
ExternalFOICe.
B----+
)
These are forces oul~de the system
over which thesyslem's 'mana99'"
ex have no control. They may be physical
or Inftuences from !he behavlour of
nature. The link has a + or • sign
Fig. 2.2 The conventions of influence diagrams.

Signs for Links
Positive Unks
If j- -+ t and !- -'- i
then the link ~ POSITIVE
Negattve Unks
then the link ~ NEGATIVE
Placing of SIgns
The sign for a Ink must be althe head or po int of the a rrow.
It ca n be on el1hef side. but must be placed so thal the link
to which It applies Is una mbiguous.
Signsfor Loops
If the loop has on EVEN number ot - signs(0 ~ on even num ber)

the the loo p is POSITIVE
~rGA~'f has on ODD number of • signs then the loop is
Fig. 2.3 Signs for links and loops.
the rate of sales, might fluctuate quite sharply from day to day and
one might want to smooth out that variation by taking the average
sales rate over a period of time, such as a month. Thus, the variable influ-
ences its own average , but requires a parameter of the period over which
the average is taken. Figure 2.3 summarizes the sign rules for links and
loops.
THE CENTRAL HEATING SYSTEM REVISITED
Having explained the conventions, we can now explain some of the factors
in the central heating problem a little more closely.
The discrepancy, or difference, between desired temperature and actual
temperature is denoted by the Greek letter epsilon , E, with a subscript 1 to
distinguish it from other epsilons. The difference is calculated in the sense
of:
Desired Temperature - Actual Temperature
The central heating system revisited 23
so that an increase in desired temperature will increase 1::1 and an increase in

actual temperature will reduce it. If actual temperature increased so much
that it exceeded the desired temperature, then 1::1 would become negative,
that is it would have continued to decrease, and the - sign would still be
correct.
It would not be wrong to define 1:: 1 in the sense of
Actual Temperature - Desired Temperature
with the corresponding changes of sign, but it would be rather confusing, as
it would imply that 1:: 1 had to be small before the heat was turned on, rather
than the common-sense approach which says that 1::, has to be large to turn
on the heat. However, even if one did define 1:: 1 in this way, the net effect
of the signs on Loop 1 would not change and it would still be a negative
loop.
The rate of supplying heat is influenced by the discrepancy, by the 'policy '
which states how large 1:: 1 can be before the heat is turned on and, obviously,
by the heat-generating power of the system. Heat arrives in the room after
a delay due to the size of the house . The actual temperature depends, as
we have seen, on the quantity of heat and also on a parameter, the heat
to temperature conversion factor, which depends on the size of the
room.
The external temperature is shown in a box to signify that it is a driving
force over which the managers of the system have no control. Such external
forces are called exogenous factors, to distinguish them from the endog-
enous factors within the system itself. The difference between the room and
external temperatures, 1::2, determines the rate at which heat will be lost
and the heat loss represents an ungovernable force of nature, about which
the home owner can do nothing . To represent this requires another param-
eter, the Temperature Difference to Rate of Loss of Heat Conversion
Factor.
The parameter for the power of the system is obvious common sense, but
why do we need to introduce these two conversion factors? Again, common
sense supplies the answer. The amount of heat in the room must be meas-
ured in some sort of Heat Units and the rate at which heat is supplied and
lost must be measured in Heat Units/Hour. Temperatures and 1::2, on the
other hand , are measured in degrees so, to get the rate of loss of heat, one
must have a conversion factor which will be measured in (Heat Units/
Hour)/Degree. The ( ) are used here to emphasize the rate at which heat
flows under the influence of the temperature difference. The reader should
work out the units, or dimensions as they are called, for the Heat to
Temperature Conversion Factor. It should now be clear that, had these two
parameters been omitted, the model would have been incomplete, so the
ability to recognize when a parameter is needed is crucial to getting a good
influence diagram. Thinking in terms of the units involved is the key: a
process called dimensional analysis, about which we shall learn more in
Chapters 4 and 5.
Before we leave this diagram it should be noted that it contains two

negative loops. Loop 1 is drawn to be larger than Loop 2, because it is the
obvious one which attempts to control the system. As we pointed out in
connection with Fig. 1.2, any goal-seeking system must contain at least one
negative loop, and Loop 1 has all the necessary characteristics of the nega-
tive feedback loop; that is, it has both a desired and an actual state , and it
has a discrepancy between the two which leads to action intended to elimi-
nate the discrepancy.
Figure 2.1 has a second loop which is clearly negative from its signs and
which acts as drain on the system, counteracting the efforts of the home
owner to be comfortable. Loop 2 is nature's response to people 's attempts
and it is the source of the profits of utility companies. Its 'desired' state is
the external temperature, so this loop is trying to control the room to the
ambient temperature while Loop 1 is trying to control it to the owner's
desires. Such conflict between loops is often the source of interesting dy-
namic behaviour.
Having studied an influence diagram, it is now time for the reader to try
constructing one.
A PRODUCTION MANAGEMENT PROBLEM
In practical system dynamics an analysis often starts from a narrative de-

scription of a problem , furnished by a client, drawn from the analyst's notes
of meetings or suggested by the literature in academic research. Figure 2.4
is such a description of a problem in production management'. We shall use
this problem , with many additional factors, throughout this book. We shall
build our diagram in small steps, to develop skill in diagramming.
Figure 2.4 should be carefully studied before reading further.
In two places, the narrative refers to eliminating discrepancies over a
period of 4 weeks, and we must digress a little to understand that idea.
In Fig. 1.2 we saw that a negative loop must generate control action to
eliminate errors between actual and desired states. In Fig. 2.1, Loop 1 in the
central heating system does exactly that by turning on the boiler, but
the rate at which the room warms up is controlled by the fixed power of
the system. People controlling managed systems have, however, freedom
to choose how swiftly or sluggishly to respond to discrepancies, and
this can be seen by again using the powerful concepts of dimensional
analysis.
For this problem the actual and desired backlogs of washing machines,
and the discrepancy between them, must all be measured in washing ma-
1 This problem is a very simplified version of a consultancy assignment the author undertook
for a manufacturing company, though the product has been changed. The real company did
use the policies and parameter values given here . We shall revisit this problem several times
throughout this book.
The Domestic Manufacturing Company, DMC, produces washing ma-

chines for sale to major retail companies. Customers tend to order
large batches of machines, with delivery required about 6 weeks after
the order so as to match the customer's own marketing plans. Making
a mach ine is very simple, mainly involving the assembly of standard
parts, and the production process takes a short time. Building ma-
chines uses up raw materials, fresh supplies being ordered from
parts makers who deliver in about 6 weeks.
DMC have never been able to forecast the inflow of new orders and
just have to try to cope with a very unpredictable order pattern. New
orders accumulate into a backlog which the company tries to keep
down to a target level.
The Production Manager considers two factors in setting the produc-

tion rate. First , he aims to eliminate any discrepancies between actual
and target backlog over a period of 4 weeks. New orders are averaged
over a 4 week period and the backlog target is 6 weeks of this level of
orders. Second, he tries to keep up with the current order level, so to
the production rate which would eliminate the discrepancy within the
planned time is added the average order rate.
The Raw Materials Manager tries to keep raw material stock up to a

target level, ordering raw materials to eliminate any discrepancies
within a period of 4 weeks . The target level is based on smoothing
production variations over 4 weeks and aiming to have sufficient
stocks to cover 8 weeks of average production. He, too, tries to keep
up w ith current usage, so to the order rate which would eliminate the
discrepancy is added the average usage rate of raw materials.
Fig. 2.4 Narrat ive of the product ion and raw material problem.
chines, or WMS (it is very bad modelling practice to measure objects in

vague 'units' and it is far better to call the units what they are) . The
production rate must, however , be measured in washing machines/week, or
WMS/week. If we wrote :
Production Rate = Discrepancy
we should be in error because the units on the left do not match those on the
right of the =. It is evident that we must have:
Production Rate = DiscrepancyfTime to eliminate the discrepancy
for the units to balance across the = sign. The 'time to eliminate the
discrepancy' is usually called the time constant. This simple analysis intro-
duces two very important ideas.
The first is that this form of control is called proportional control:
the strength of the control action is proportional to the size of the discrep-
ancy to be controlled. It is a very common mistake in system dynamics
models to omit the time constant. Dimensional analysis shows that the
time constant is essential to the equation, so not explicitly including
the time constant in the preceding equation is tantamount to including
it with the numerical value of 1. There is no earthly reason why it has to
be 1, nor is there any reason to suppose that a value of 1 will produce the
best behaviour in the system. Omitting the time constant is, therefore,
not only a dimensional mistake, it also denies the analyst the use of
what might be an important opportunity to design better behaviour into the
system.
The second idea is that the time constant is an aspect of the system's
policies, generating action in response to information about the conse-
quences of earlier actions. It can and should be chosen to suit the system's
need for robust behaviour. In this case the managers have chosen to use
time constants of 4 weeks, which may, or may not , be good values. The
system dynamics analyst's task is to find out whether x weeks is a value
which makes the system work well and to propose another if it is not. As we
shall see, the modeller has a very large degree of freedom to redesign
systems to improve their behaviour, and it is that freedom which makes
system dynamics a practical possibility.
Reverting to the problem, it usually requ ires a little care before attempt-
ing to draw a diagram from a narrative. A good method is to highlight in
the narrative the names of the variables which will be needed, to pin-
point parameters and, above all, to identify the flows as an essential
precursor to the need to think physics. Figure 2.5 reproduces Fig. 2.4, but
with items noted in italics. This process is called 'parsing', by analogy with
grammar.
In practice, parsing is usually more difficult than Fig. 2.5 implies. Figure
2.4 was a completed narrative for a textbook problem, but practical
narratives are rarely complete. The attempt to parse reveals the incom-
pletenesses in the narrative and thus takes the analyst back to the problem
which has not been understood if the narrative is imperfect. This is a good
example of what was meant in Chapter 1 about the methodology of system
dynamics helping one to understand the problem more clearly.
Clearly, what flows is washing machines, but there seem to be two sub-
flows: orders and raw materials. One of the tricks of system dynamics is not
to get bogged down in too much detail in the early stages of an analysis. In
this case, the narrative gives no reason to worry about the difference be-
tween, say, motors and paint as raw materials, and we can assume that the
raw material manager knows his job and orders raw materials in 'washing
The Domestic Manufacturing Company, DMC, produces washing ma-

chines for sale to major retail companies. Customers tend to order
large batches of machines, with delivery required about 6 weeks after
the order so as to match the customer's own marketing plans. Making
a machine is very simple, mainly involving the assembly of standard
parts, and the production process takes a short time. Building ma-
chines uses up raw materials, fresh supplies being ordered from
parts makers who deliver in about 6 weeks.
DMC have never been able to forecast the inflow of new orders and
just have to try to cope with a very unpredictable order pattern. New
orders accumulate into a backlog which the company tries to keep
down to a target level.
The Production Manager considers two factors in setting the produc-

tion rate. First , he aims to eliminate any discrepancies between actual
and target backlog over a period of 4 weeks. New orders are averaged
over a 4 week period and the backlog target is 6 weeks of this level of
orders. Second , he tries to keep up with the current order level, so to
the production rate which would eliminate the discrepancy within the
planned time is added the average order rate.
The Raw Materials Manager tries to keep raw material stock up to a

target level, ordering raw materials to eliminate any discrepancies
within a period of 4 weeks. The target level is based on smoothing
production variations over 4 weeks and aiming to have sufficient
stocks to cover 8 weeks of average production. He, too, tries to keep
up with current usage, so to the order rate which would eliminate the
discrepancy is added the average usage rate of raw materials.
Fig. 2.5 Parsed narrative of the production and raw material problem.
machine equivalents'. Our aim is to analyse the robustness of the overall

policies for running the business, not to do detailed order planning.
Secondly, we note a significant delay in obtaining new materials, but not
in the production process. This tells us that we need to distinguish between
the ordering and the arrival of raw materials, but not between the rate of
starting production and the rate of shipping from the factory to the cus-
tomer. Finally, the policies for controlling production and raw material
ordering are described and the parameters identified.
The first step in drawing the diagram is to identify the physical flows, and
the reader should pause and attempt to draw them before going further.
The physics is shown in Fig. 2.6, the rather strange layout of which is to
give consistency with the final diagram in Fig. 2.7. In Fig. 2.6 New Order
Inflow Rate is drawn in a box to show that it is an unpredictable driving
force, as was external temperature in the central heating diagram. Orders
accumulate into a backlog, which is depleted by production. The production
rate depletes raw material stock which is fed by raw material arrivals, in
tum a delayed version of raw material orders.
Figure 2.6 is the skeleton of the system's consequences on to which we
must impose the nerves and muscles of the information and action compo-
nents. To emphasize that, Production Rate and Raw Material Order Rate
have been enclosed in ellipses. These are not part of the influence diagram,
but are temporary additions to draw attention to the fact that these two
variables are the controlling flows for this system to which the information
and action parts of the influence diagram should lead. Recalling the analogy
of water flowing in a set of pipes and tanks , these are the valves which
regulate flow. Again, the reader should try drawing the regulating influ-
ences before looking at Fig. 2.7.
Figure 2.7 shows the complete diagram. It is unlikely that the reader used
these variable names, though they are not important in themselves . It is also
unlikely that the reader's solution was as clear as this and involved no lines
crossing. Laying out a diagram so that it is easy on the eye is important to
give the impression of a smooth flow of control. The reader now has to go
back to Fig. 2.5 and verify from the narrative that the factors in the diagram
are those mentioned there and that the signs on the diagram are those
required by the narrative.
For example, the narrative comment that the average usage rate (of raw
materials) is added to the raw material orders required to eliminate the
discrepancy requires that there be a + sign on the direct link from Average
Production Rate to Raw Material Order Rate.
The pattern of signs on the links from Desired Backlog and Backlog to
Backlog Discrepancy should be compared with those on the links from
Desired Raw Material Stocks and Raw Material Stocks to Raw Material
Stock Discrepancy. The Backlog Discrepancy takes the form:
Backlog - Desired Backlog
while the Raw Material Discrepancy takes the form:
Desired Raw Material Stock - Raw Material Stock
Again, the narrative makes this clear. The production task is to keep Back-
log down to a level which will satisfy the customers, while the Raw Material
Manager tries to keep Raw Materials Stock up to a level which will ensure
that the factory can keep running . Great care is needed in ensuring that
discrepancies are properly defined, but the influence diagram provides a
vital clue. The influence of New Order Inflow Rate on Backlog is positive,
that is it will tend to drive Backlog up and the manager has to try to bring
W
_ Ra M aterial
»>
.---- s tocks
++
New Order
Inflow Rate
Raw Material
Arrival Rate
+
.
Fig. 2.6 The physics of the production and raw material problem.
Averaging Period
- - _ lex Produclion /la/e
Averaging Period
IOf Old« Rat. ........
./ - - - <,
....... .... Average
/ "Production
~ / \ Rote" <,
Average
Order
"+
• Rote +~ ".. Desired <,
\ / + Row Material ....
/ , +
--
Stocks" .....
/ ' . Desired \ / /' \ \
Backlog
/ of Orders - - \ / // \ \
/
.,. \_-
..
I / W""'atAve~ge
PlOduclion In
*+
'f
\
/ Backlog o.sired Raw Row Material \
/ Discrepancy " I / ldalMial Slocks - Stock Discrepancy
/ +~ ."." \ I / /' ~ -, \
I W.e'" at Average
Orders as o..,ed /, /'/' '\ \
\ Backlog .\ \
I I NegatIve" I / / _ _
MorenOI I~OW .". - "
\ I \ Loop 1, .....
\
+' +too/ _»> "':;": ,'"
\ prag:f:
: .s- \
\ Negative
Backlog of
\ + Orders I
I I : Loop 2
,
\ I It-
I \. / \
riffle 10 Correcl Row Moteriol ",
Order Backlog Arrival Rote \
Discrepancy ....c-
+ \
Time10 Co"ecl
Raw Ma/erial
Discrepancy
Fig.2.7 The complete production and raw material problem.

The list extension method 31
it back down. On the other hand , the influence of Production Rate on Raw
Material Stock is to drive it down, and the manager has to try to keep it up.
Finally, this diagram has only two loops. The reader should prove that
starting, for example , at Production Rate and moving to Average Produc-
tion Rate does not lead back to the starting point.
Both of the loops are negative, goal-seeking, mechanisms, as is shown by
the pattern of signs. That is not a sufficient guarantee that they should be
negative and we must apply the qualitative test of reasoning about what the
loops are trying to do. Loop 1, for example, represents the Production
Manager's efforts to keep Backlog to a satisfactory level. It should, there-
fore, be negative and it is. This all adds to one's confidence that the influ-
ence diagram is correct. This process of building one's confidence in the
model by seeing it pass more and more tests is the validation which was
mentioned in Chapter 1.
Whether, of course , these loops are successful in achieving what they are
supposed to achieve is an entirely different matter, and we shall find later
that they are far from being good at their jobs, because of poor design.
Influence diagrams are a vital tool in system dynamics and they usually
look very obvious after they have been built. The newcomer to system
dynamics, however, faces the problem of learning how to get started on an
influence diagram for a new problem, so we now turn to three methods for
building diagrams.
THE LIST EXTENSION METHOD
An influence diagram is a list of factors in a problem, together with arrows

and signs showing the relationships between them . List extension is based
on the rather obvious idea of starting with a small list and gradually extend-
ing it until a diagram emerges . This concept is shown in Fig. 2.8 and the
reader should use a piece of paper to cover up all of that diagram except the
column headings and the extreme right-hand column. The paper can be
moved to reveal more of the diagram as the explanation proceeds.
The column headed 'Model list' is the starting point for developing a
diagram. It contains the name of a variable which it is the purpose of the
model to explain or for the control of which policies are to be designed. In
this case, we shall use the Temperature Discrepancy for the central heating
diagram. In a practical case, one can use two or three variables, but if more
than that are put into the model list, the purpose of the analysis is probably
confused and one should revisit Stage 1 in Fig. 1.3.
This list of one item is now extended by writing in the column headed
'First Extension' the names of the variables which directly and immediately
influence the variable in the model list, adding the appropriate arrows and
signs. It is essential to be clear in one's mind about direct influences; in this
I
1
Fifth Fourth Third Second ,,,

First
Extension Extension Model
Extension Extension
,, Extension List
,,
,
------------~----- ------------:-- ----- -- - - --------~-----------------~- ------ -- - -----~----------_ .
: I : I
, I 1 Heat to Temperature :
: : _ _ __ : Conversion Foctor I
System I I ,... ..... - --_ I " I

'PolIcy' .............. I '...... -_1 -,
I /......, i"....... ,.I
<, I / " I .............. "-
<, I /" I I....... I "
<,, + /' I I ....... I "
........ JI./: I............. : ~
Rate of Supplying II) +~ Rate of A r r i v a l : ""'.....' Actual Temperature
/' Heat HS of Heat In Room ~ -, I
' / I 1 ..... I 1"-
+ In Room
,....... I I '- I /~ -,
......., I <, I./' I "-
// 1 I , ~/ I -,
Powelo{ / I I ,+ / <, I ~
the System I I , // : .......... ' -, -
: , Q u a n t i t y of I ..... I '"
I ' Hea t In I 'Temperature
J.I - - _
, ~ Room ' I - - -:- Discrepancy
, : - I I I "+ e,
Temperatule I I 1/
Difference to
Rate 0{ LOS$ of __ ~ ~RateHoflt°slng : / /"
I eo I / ,/
Heat I / "t I /' //
Conversion : / + /1-' /
FactOI , / .--
I / ....... / Desired Temperature
,/ / on Thermostat
,/ /
I / ,/
)/ .>
/ I "
Difference / I " /
External Between : ...... "

Temperature --;----~Internaland
External ~...!
+' ---
_--- : 1
Temperatures I I
£Z: :
Fig. 2.8 List extension for the central heating problem.
case, the temperature discrepancy is, by definition, the difference between
target and actual temperatures in the room. One might be tempted to
write 'External Temperature' in the first extension, but that is an indirect
influence.
The second extension continues the process by writing down the direct
and immediate influences on the variables in the first extens ion. A chain of
influence terminates when a parameter or an external force is encountered.
As each extension unfolds, a check has to be made to previous extensions
for links from earlier variables. Thus, in Fig. 2.8 we see links running from
right to left which emerge as extensions are built. Notice , however, that Fig.
2.8 is nothing like as neat and orderly as Fig. 2.1; in particular, one of the
loops is in the shape of a figure of eight. Diagrams built using list extension
almost always have to be drawn again to minimize crossing links and por-
tray a smooth flow of influences and clearly visible feedback loops.
The reader should now repeat the exercise for the central heating prob-
lem, starting from any other variable, and repeat it again for the production
management problem. In each case, the process should lead automatically
to the same solution each time. It is one of the very attractive aspects of list
extension that two analysts tackling the same problem should come up with
recognizably similar solutions even if they started with different ideas as to
what were the key variables . This is a solution to the 'two-analyst problem',
and it is vital to good modelling that one should get the same end view of
the system, regardless of one's initial prejudices about significant variables .
List extension is a good method for starting work on a problem, espe-
cially in one 's early stages of developing skill in influence diagramming. It is
usually not necessary to go past four or five extensions as the pattern of the
diagram will probably have emerged by then and one can use one of the
other two methods to proceed. It is, however, essential to have the patience
to redraw the diagram to make it look neat.
As a rough guide, the first stage of modelling a new problem should
not involve more than about 15 or 20 variables and parameters. It is very
easy to expand a model in later stages of a project and it is better to start
with something simple. The key to successful modelling is to keep one's
understanding of the model and what it says about the problem ahead of its
size. A model which is too large to start with restricts that development of
understanding and insight which was emphasized in Fig. 1.4.
THE ENTITY/STATEfTRANSITION METHOD
The emphasis on the flows of something in a system and the analogy of

someone attempting to regulate the flow of water in a set of pipes and tanks
by turning valves makes the idea of identifying all the entities in a system
and mapping their transitions from one state to another a very effective
approach to problems in which the physics are more complicated than in the
central heating or production management cases. We shall first describe a

very simple problem in manpower planning and then show how this method
can lead very easily to an influence diagram .
Consider the following narrative account.
A firm of management consultants recruits new staff to fill any shortfall
between the number of qualified consultants available for projects and
the number needed to cope with the demand for the firm's services.
Recruits undergo a protracted training period before they become
fully qualified consultants. When they complete their training, they
join the consultancy team. Qualified consultants tend to stay with the
firm for a period of years before they leave to go into industry as
managers . Some of the qualified consultants employed by the firm have
to be used to supervise the trainees, on the scale of 1 supervisor to 10
trainees, and these supervisors are not available for fee-earning assign-
ments . The trainees are not able to work on assignments. The total
number of qualified working consultants the firm needs is driven by the
demand for consultancy services.
The first step is for the reader to parse this statement, as was done earlier
for the washing machine problem. Notice that the last sentence repeats
information from the first sentence. Such duplication is very common in
practical modelling and the analyst has to be careful to avoid double-
counting the same factors.
Having identified the factors in the problem by parsing, we move to the
entity/state/transition method, the steps in which are listed in Fig. 2.9.
In this problem, the entities are Trainees and Consultants. Although
individual trainees become consultants in due course, two separate groups
of people can be identified at anyone time.
A state is a condition in which an entity remains until something causes
a change. By the water flow analogy, water in a tank would remain in that
state until a valve is turned to allow the water to flow to another tank . For
this problem, trainees can only be in the state of undergoing training and
consultants can only be in the state of being available either to work on
projects or train new recruits.
The flows cause the numbers of a particular entity in a given state to
increase or decrease. For the trainees, the direct and immediate influences
are the inflow of new recruits and the outflow at the end of the training
period. For the consultants, the inflow is the rate of completion of training,
at which point individual people move from one state to another, and the
outflow is the rate at which qualified consultants leave the company. Thus ,
as step 4 enjoins, we have identified a flow which interconnects the two
entities. Step 4 also stipulates that one should be alert for delays, and the
parsed narrative makes it clear that there are two delays: the training of
recruits and the period for which qualified consultants stay with the firm.
Adding these delays to the picture gives the diagram shown in Fig. 2.10.
1. identify all the separate entities or actors in the problem.
2. For each entity identify all the possible states in which members of
that entity can be.
3. For each state, identify the flows which can cause the state to
increase or decrease.
4. Check for connections between flows. Does the outflow from one
state feed another?
Ensure that any delays in flows are represented, especially when

relating the outflow from one state to the inflow to another.
5. Identify the controlling flow rates which drive the system. In gen-
eral, these will have arrows coming out of them, showing that they
influence something, but no arrows going in to show what influ-
ences them.
6. Identify and represent the information and action influences on the

controlling flow rates. This is usually done from the parsing of the
narrative account.
Fig. 2.9 The steps in the entit y/state tran sition method.
r
+ Trainee Qualified
+ Consultants
T~~~2J ">.
c onSl,Jltants-
, (
~~~~~~~fsro~
-
training duties '\
+ Tralnin~ + Consultant
Trainee
Recruitment
Rate
DniAti • Completion
Rate
DIM'\.OV • Leaving
Rate
Fig. 2.10 The entities, states and trans itions in the consultant problem.
It should be clear that the entity/state/transition method is geared to the

injunction to think physics and the notion of ident ifying the direct and
immediate influences is that used in list extension. It is not surprising that
the various methods of building influence diagrams have much in common;
they are directed towards the same end .
In Step 5, we identify the controlling flow rate as the trainee recruitment
rat e and must now move on to the information and action links to complete
the diagram . This is usually rather more difficult and it is often useful to do
a minor list extension exercise to get one 's ideas straight.
The narrative states that the number of consultants required for work
depends on the demand for the firm's services. It also states that some of the
qualified consultants have to be used to train recruits and we shall assume
for the moment, because the narrative does not tell us to assume anything
else, that qualified consultants can be moved instantly between working on
assignments and training recruits. This leads to the final picture , shown in
Fig. 2.11. Note the parameter, 'Time to Correct Consultant Discrepancy',
which was not mentioned in the narrative but which, for reasons to be
discussed later, has to be introduced into the model. The reader should
study Fig. 2.11 carefully, relating it to the narrative for this problem and
comparing it to the previous examples, noting and understanding the simi-
larities and differences.
The point made in the previous paragraph about not making an unsup-
ported assumption is rather important. System dynamics models, as we shall
see when we come to write equations, are very easy to write. That means
that they are also easy to extend as more information comes to hand and as
understanding of the problem deepens. However, the ease of writing also
tempts one to make the model larger than it needs to be in the early stages
of a project. It is essential to keep one 's understanding of a model ahead of
its size and that means that it is good practice to keep the model simple to
start with and only to make it more complicated in due time. Making a
model more complicated is not the same as improving the sophistication
with which it represents a problem, and subtlety is always to be preferred to
mere size.
THE COMMON MODULES METHOD
Figures 2.1 and 2.7 use the same unit of structure in that they both contain
a flow which, after a delay, leads to an effect. In one case the flow starts with
the supply of heat, in the other with the ordering of raw materials, but, apart
from these differences in names, the two portions of the influence diagrams
are identical. Figures 2.7 and 2.11 both contain information and action
components in which target and actual states are compared and action is
taken to correct a discrepancy over a time constant. Figure 2.11 uses the
same format for the delay in training recruits and for the duration of
employment of qualified people . In short, managed systems contain com-
mon modules of structure which recur in utterly different problems. Once
one has learned to recognize these modules , building influence diagrams
becomes quite easy. Caution must be exercised , however, against simply
throwing modules into a diagram without taking sufficient care to ensure
that they are the right ones for the particular case.
Some of the standard modules are shown in Fig. 2.12, which uses 'level'
to suggest the analogy of the contents of a water tank . The reader should
not bother to remember the names of the modules, but should learn to
recognize the phenomena they represent.
Required
Number ot + Demand for
------ - - ------ - ------ -- - - Qualified .. - - - - - - C onsulta ncy
Consultants Services
Available for
+/ " Projects
~
Qualified __ - - - - - - - - -- -- - - - _ "-, -,
Consultan t - - <, -,
Discrepancy - <, -,
-, -, Qual/fled
\ Consultants
per Unit
\ of Demand
Qualified Qualified
I
Ra"oof Consultants - Consultants
Supervlsors - - - - .. Required to - - - - - - - - - - ~ Available
10 Trainees / .. Supervise for Proje cts
Trainees
/ + •
I + I
I I
I I
nme to I I 'I
Correcl I I
Consuffonl + Trainee + Qualified
Dlscr~pancy:I .:">. Trained

(~~7i~U~~~~r ~ a ssignments or
\ training dutie s
\ I
\ \
\" +\ nt
Tra ine e Leaving
Tralnln~ DIM'lDV + •Consulta
..... • Recruitment Dm:v. le tron Rate
+• Comp
Rate Rate
Fig. 2.11 The completed consultant problem.

+ -
Inflow---+- Level ~ Drain
A - SImple Inflow/Ouffiow
B - Delayed Inflow
level I ~ Tr~~:;cn ~level2
C - Transitional Flow
D - Delayed (Irreverslble) Transitional Flow
_{~z;;-~*--+·~~:~t: \+
Levell Level2
+\ Inflow Outflow}-
' - - - 10 • from ---'"
Levell + Leve12
E - Delayed (Reverslble) Translflonal Flow
Ta rg et Sta te _ ......... +
Actual state- ,/
.-
Discrepan cy....
\
\",
Time Constant
, Variable +~S~~~d
/
11
+t ¥ /
Flow Role smoothing
F - Proportional Control G - ~lhed Informaflon
Variable 1
(eg Prlce) - -
m
- ,,- -
+ 0(- Vartoble 2
-~ (eg Demond)
H - Behavioural Transformation
Fig. 2.U Some common modules.
Case A, simple inflow/outflow, occurs in Fig. 2.6 when orders flow into
backlog and production depletes it. Case B, delayed inflow, represents the
ordering of raw materials in Fig. 2.6, except that the contents of the delay,
the amount of raw material ordered but not yet delivered , is not shown
in that illustration, the reason being that the narrative account did not
mention it. Clearly, the contents of a delay are present in a system and
Which is the right method to use? 39
it is typical, in practical work, that narrative accounts do not mention
everything. The value of the common modules is that they help one to
recognize items that may have been omitted from a narrative. The reader
should now study the earlier diagrams in this chapter and find examples of
cases C, F and G, noting how modules sometimes overlap. Figure 2.7 uses
cases F and G combined to represent the influences on production rate.
Case C implies that the transition can be in either direction, such as
swings of allegiance between one political party and its rival ; a positive
transition rate flowing from Levell to Level 2, and a negative rate flowing
the other way. In other cases, such as the flow through a pipe, or the
ordering of raw materials in Fig. 2.6, the flow cannot be reversed, and this
is shown in case D. If, for example, the qualified consultants required a
period of time to adapt from working on projects to training recruits, Case
D would be an appropriate module, and the delay contents would be those
consultants who were neither available for work nor to supervise training.
Unfortunately, case D would not allow them to move back to project work
as needs changed, and for that we should have to use Case E, which
provides for a reversible flow, the two delay contents not being shown;
readers should practice drawing them for themselves',
Case H represents a common phenomenon in managed systems whereby
one variable is transformed into another, sometimes by managerial choice,
but also, as shown here, by an ungovernable force of the economy in which
price is transformed into demand. The Greek letter e, phi , denotes that
there is a function or formula which models the transformation of one
variable into another. The effect may be + or - or even +/-, meaning that
in some circumstances it is + while in others it is -. This is an example of
non-linear behaviour in managed systems, which we shall study in much
more detail later in the book.
WHICH IS THE RIGHT METHOD TO USE?
We have studied three methods of building influence diagrams so it is a

natural question to ask which method to use in different circumstances. In
practice, there is no simple answer, though we shall use this section to
suggest some broad guidelines. The reader should try to apply them to the
influence diagram case studies in the next chapter, as that will help you to
see how those models came about.
It should be realized that an experienced system dynamics modeller will
often use none of the methods formally, in the sense of actually writing out
a list extension or the entities, states and transitions. The modeller will,
2 In quantitative modelling, it would be necessary to ensure that the flows were not working in
both directions at the same time, a problem we shall address in a later chapter.
however, be using these methods, especially the common modules , though

perhaps without recognizing their names, mentally as the diagram pro-
ceeds. The diagram seems to emerge on paper without effort , though there
is a good deal of thought behind it.
For the less experienced modeller, list extension is usually a good way to
start, mainly because it requires one to stipulate the model variables, which
is usually a good way of being clear about the purpose of the model. The list
extension should, however, be stopped after about four or five extensions
and the diagram drawn again to be as clear as possible.
When a diagram drawn using list extension is reviewed it typically leaves
the analyst with a feeling that he is being rather vague about some aspects
of the problem. This is where the entity-based method comes in to resolve
ambiguities about what is flowing and how it flows. Being clear about the
flows helps immeasurably when it comes to mapping the information/action
elements which control the flows.
The common modules method is a useful check on the correctness of the
representation of the flows.
SOME GENERAL GUIDELINES FOR DRAWING

INFLUENCE DIAGRAMS
Having studied some examples of influence diagrams and considered three

approaches to their construction it is time to offer some general comments
on influence diagramming in practice, as shown in Fig. 2.13.
These guidelines are not absolutes, and, as we are about to see, the strict
rules of influence diagrams, shown in Figs. 2.2 and 2.3, should not be made
into a handicap to thought. Readers should, however, review Fig. 2.13 from
time to time and develop their own guidelines as practice is gained and skill
increases.
DIAGRAMS AT A MORE AGGREGATED LEVEL
A fully detailed influence diagram can sometimes be an obstacle to commu-

nication with a sponsor and it is often useful to draw a diagram to show less
detail while still capturing the main features of the problem. For instance,
Fig. 2.14 is a rather generalized portrayal of the relationships between an
international mining company , 'Anglo-Consolidated Zinc' (ACZ), and its
operating subsidiaries around the world.
ACZ derives cash flow from the distributed profits of the mines of which
it is part owner. Money retained by operating companies is invested either
to maintain an existing mine in production or to develop new mines. ACZ
invest their cash flow either in exploration to discover new reserves or to
develop new mines. The diagram shows these as the key features of the
Diagrams at a more aggregated level 41
1. Always be clear about the purpose for which the model is being
built. Is it a top-level, strategic view or a more detailed, tactical,
problem? Who is the customer for the model - a fee-paying client
or the academic community?
2. Always be alert for the difference between the information/action/

consequences parts of the paradigm.
3. Choose variables which are defined so that they can, in principle,

be measured and which can vary over time. 'Attitude of Custom-
ers' would a poor name, 'Customer Satisfaction' would be a good
one.
4. Whenever a va riable such as 'actual backlog ' is seen, look for the
corresponding 'desired backlog' and vice versa. As we have seen,
without the comparison of an actual and a desired state, negative,
goal-seeking, feedback does not exist and control cannot be
achieved.
5. Distinguish carefully between actual values and perceived values .

For example, actual 'Customer Satisfaction' can almost certainly
not be known . The best that could be achieved would be a percep-
tion of it, so 'Perceived Customer Satisfaction' would enter the
diagram in addition to the actual value. Perceptions usually lag
behind reality, so the time delay would need to be included.
6. Watch out for unintended effects. An increase in the production

rate might lead to a reduction in quality, w ith consequences, per-
haps, for real and perceived Customer Satisfaction.
7. If a link between two variables cannot be easily explained to others

it may be because an intermediate factor needs to be included. A
direct link from production rate to customer satisfaction might be
hard to understand without including the effect of production on
product quality.
8. Above all, keep the diagram simple to start with.
Fig. 2.13 Some guidelines for drawing influence diagrams.
problem, though without the use of solid and dotted lines, and without
identifying parameters. The exogenous driving force of metal prices, which
even ACZ is not big enough to influence, is shown with the usual box. In
short, this is a top-level view of the problem. It is, in fact, deliberately drawn
to be a rather poor top-level view, because it says little about the policies by
exploration D + Disc
over ad
+ Spending 1 • Reserves
%O~FIOW . I J-
Devotedto~
exploration
/
Spending +
"InNewu·~
"L.--t--- ~
I
D
'T~\~ /
+ +"00
:Tl~ 2
~
(
ACZ's Cash Depreciation Profits
Flow . / \ \
+ Mine
Cash
y '---- -
Inveslmentln
Exis~ng Mines
Flow
+~~roooo+ Profit
_ _ _ Distrlbut"",--+~
Profits
Fig.2.14 Salient featur es of 'Anglo-Consolidated Zinc' model.
which ACZ decides to split its cash flow between the two streams other than
to imply that , when ACZ has sufficient discovered reserves, more money
will be put into developing new mines and vice versa. The ideas of 'target'
or 'acceptable' reserves and mine capacity are not present. The diagram
says nothing at all about the operating companies ' policies.
Oth er apparent faults are that the diagram includes delayed flow modules
of type D from Fig. 2.12, but it shows only the inflow to the delay, Explora-
tion spending measured in $/year, no outflow, such as Exploration spending
becoming effective, and Discovered reserves must be measured in tons
of metal, so there is a dimensional inconsistency, caused by a missing
parameter.
In short, the diagram seems to break all the rules of influence diagrams
which has been done to engender in the reader the habit of looking critically
at an influence diagram: a most essential skill to develop. The aggregated
diagram may, nevertheless , be a more useful diagram for discussion with
senior managers in ACZ. Rules are often said to be for the observance of
fools and the guidance of wise men. That is putting it a little too strongly,
but the reader should break the rules when it helps, especially when some
practice has been gained in following them.
It would probably be good practice for the reader to redraw Fig. 2.14 in
The cone of influence diagrams 43
_ - - - D i s c overed -
Emphasis
..
_~ the F~Uffi
on exp lo ration ::----........-
'~~~)+
~_ _.....:+:..rc~p~~:~n
In Operating
~------ Mines
Fig.2.15 A very aggregated view of 'Anglo-Consolidated Zinc'.
more detail, observing the modules, introducing parameters and distin-

guishing between information, action and consequences by using solid and
broken lines.
Figure 2.15 is an even more aggregated view of the same problem. It
introduces two new variables , Security for the Future and Stability for the
Present, to capture what is really conferred by the development of reserves
and investment in mine production capacity . The effect of prices is shown as
an exogenous factor , with a + sign, and the percentage retention of profits
by the operating comp anies, which would form part of the financial agree-
ment with the nation where the mine is located, is strongly emphasized.
Figure 2.15 is, perhaps, the most appropriate simplification of all to convey
ideas to the most senior management of ACZ, not because they are stupid,
but because it corresponds most closely to the problems with which they
must deal at a strategic level (this broad view is closest to the idea of a
'causal loop diagram').
Readers should study these two diagrams quite closely and form their
own views on how well, or badly, they conform to the guidelines discussed
above. Neither diagram is perfect.
THE CONE OF INFLUENCE DIAGRAMS
We have drawn two different diagrams for the same problem and this leads
to the ideas shown in Fig. 2.16, namely that it is not only legitimate, but it
can also be desirable to draw many diagrams of the same problem.
Sponsor's Analyst's
Pointof Point of
View View
Fundamental Most
Insight Significant
1'lesu~s
Main
Srmptoms Features
o Concern of Problem
Main Most
Operations Variables
Descnbed
Fine 4
:-----L J-- --H-+-- _
Computer
Detail
- - --- - - - -- - - - - Coding
Level
Fig. 2.16 The 'cone of influence diagrams.
Each of the ellipses in Fig. 2.16 symbolizes an influence diagram, the size
of the ellipse suggesting the amount of detail the diagram contains. The
diagonal ellipse implies a diagram in which some parts of the problem are
shown in great detail , while other parts are shown in broad outline. Notice
the division between the sponsor's and the analyst's points of view. Clearly,
there is no magic about having four levels, and it is usually fruitless to argue
about whether a given diagram is at a particular level. The true point is that
influence diagrams are flexible tools and should be adapted to the purpose.
Notice, though, that the expression 'Conceptual consistency, variable
detail' in Fig. 2.16 does not mean that the same variable names have to be
used in diagrams at different levels. As we saw in Fig. 2.15, new variables
representing security and stability were introduced because they were ap-
propriate for the level 1 view. These variables do not exist as such in Fig.
2.14 because they would not be appropriate at that level.
In a practical project, the problem is usually first encountered at level 2.
Further analysis leads one to level 3 and perhaps to level 4 if a simulation
model is constructed. The emphasis in Fig. 1.4 on insight and understanding
implies that one might then work back up the levels to convey the most
significant results. The level 1 diagram is often drawn only at the end of the
study.
Level/rate diagrams 45
LIMITATIONS OF INFLUENCE DIAGRAMS
As we shall see in later chapters when we deal with quantitative models,

even the level 4 diagram does not show every single detail of the equations.
Attempts have been made over the years to adapt the influence diagram
conventions to include details of such factors as choices between two alter-
natives, non-linear effects and so on . These attempts have not succeeded
because they destroy the essential value of the influence diagram, namely its
simplicity and its value as tool for communication which does not require an
intimate knowledge of the simulation techniques of system dynamics.
LEVEL/RATE DIAGRAMS
Before discussing what constitutes a 'good' influence diagram and summa-

rizing this chapter it is necessary to mention another form of diagram which
is used in system dynamics. This uses standard symbols for the variables
which correspond, on the water flow analogy, to the levels in the tanks and
the flows in the pipes. A third symbol is used for so-called auxiliary vari-
ables, which are neither levels nor flows. Figure 2.17 shows part of Fig. 2.7
drawn in this way.
These diagrams make very clear the distinction between what flows and
where it accumulates. The 'cloud' symbols at the ends of the line represent
sources and sinks from which flows arise and into which they vanish when
they are no longer of interest, though if Fig. 2.17 had been drawn in full in
<,
~" " \
\
\
\
\
\
\
\
Backlog of
OrdElB
1
New Order I
Inflow Rafe I
I
\
\
\
\
\
, I
-,
--_/ /
Fig.2.17 A partial level/rate diagram.

1. Have the purpose and the target audience for the diagram been
carefully chosen? It would be no good using a level 4 diagram for
a level 1 audience.
2. Are the factors which it includes consistent with the purpose?
3. The objective of system dynamics is policy analysis, so are the

policies clearly shown in the diagram?
4. System dynamics also aims to produce policies which are robust

against a range of circumstances, so are the exogenous factors
which might present the system with setbacks or opportunities
clearly identified?
5. Are the variables capable of being easily explained to the target

audience, are they capable in principle of being measured and can
they vary over time? A variable such as 'Security for the Future' in
Fig. 2.15 might be measured in tons of metal in known reserves,
though what those tons provide is security, in which a top-level
audience might be more interested .
6. If it is a level 1 diagram, does it capture the most significant

insights and was it drawn after careful analysis?
7. Can the diagram be redrawn to be one level lower or higher

without losing conceptual consistency?
8. Has the diagram been constrained by too slavish an adherence to

the conventions? If it is too complicated by solid and broken lines
it may fail to communicate. On the other hand, the conventions
ident ify important aspects of a system. A difficult choice may be
needed.
9. Is the diagram neat and tidy, with a minimum number of lines

crossing? Can feedback loops be clearly seen?
10. Do I understand the diagram myself before I try to explain it to

someone else or to write about it?
Fig. 2.18 Some criteria for assessing influence diagrams.
this style, the arrow would have led to another box for Raw Material Stock.
The circle denotes the auxiliary variable and other symbols exist for delays.
Obviously, such diagrams are incredibly tedious to draw by hand or with
a graphics package, but they are the basis on which two of the software
Summary 47
packages for system dynamics work. In those packages, symbols are drawn
on the screen using standard icons, which reduces the labour involved.
Once the symbols have been drawn, some of the equations for the simula-
tion model can be generated automatically, though the user has to create
the rest. Equivalent facilities exist in other packages which use influence
diagrams, rather than level/rate diagrams. The topic of software is explored
in more detail in Appendix A.
WHAT IS A 'GOOD' DIAGRAM?
Simply drawing diagrams is not enough: one must know how to recognize a
good diagram from a poor one. There are no simple answers, but some
criteria are listed in Fig. 2.18. These should be reviewed carefully and
applied to the diagrams which appear in this book.
SUMMARY
This has necessarily been quite a long chapter, as it sets out to explain the
foundation on which system dynamics analysis is based. We have studied
the development of diagrams and suggested conventions. From that point,
a number of diagrams were developed.
Towards the end of the chapter, we introduced the important notion of
drawing different diagrams for the same problem and put forward some
criteria for a 'good ' diagram.
The fundamental point is that the diagram should portray the interaction
of cause and effect within a system, because it is that interplay which will
govern its ability to generate its own future behaviour.
In the next chapter, we shall study some more diagrams, some of which
will be the basis for the quantitative models developed later in the book.
CHAPTER 3
Influence diagram
case studies
INTRODUCTION
In the last chapter we studied some simple examples of influence diagrams

and developed conventions for drawing them. We now need to develop skill
in handling rather more realistic problems.
The cases discussed in this chapter are intended to show the practically
limitless range of problems to which influence diagrams, and system dynam-
ics in general, can be applied. These cases deal with manpower planning,
product development, a predator/prey system, military combat and the
decline and fall of a civilization. This is a wide range of subjects, but many
years of experience suggest that analysts are restricted only by their ability
to grasp a problem and see how it can be framed in system dynamics terms ,
and it is that insight which these cases are intended to foster .
Each problem is introduced and analysed in turn . Most of them are
revisited at the end of the chapter in what amount to exercises for the
reader to extend the diagram from a further narrative description.
These problems are much more complicated than those in the previous
chapter and there is, unfortunately, no substitute for carefully going
through them and trying to solve them oneself before looking at the 'an-
swer'. Bear in mind that, when studying managed systems, there is rarely an
unambiguously 'right' answer. Two reasonably competent analysts should ,
however, come to recognizably similar answers, though there might be
differences of interpretation. Do not , therefore, look upon the solutions in
this chapter as being definitive.
The solution to the first problem will be explained in a good deal of detail
to give the reader some ideas on how to approach the other case studies.
First, we must consider the vital question of the purpose of a model.
THE PURPOSE OF A MODEL
The essential first step in studying any problem is to be clear about what the
model is for, as was mentioned in Fig. 2.18. A model's purpose has a strong
influence on what it includes, so being clear about the purpose is an impor-
tant step in good analysis.
The consulting firm 's problem 49
There are three main aspects of purpose:

• What has happened to the system to make its managers worry about it?
This is usually expressed as a pattern of behaviour, experienced in the
past or expected in the future, which is in some way unacceptable. The
pattern is often referred to as the reference mode.
• What is the time period of concern? If the problem concerns corporate
growth over a period of years there is little point in representing pro-
cesses involving delays of days.
• What are the main policy levers which managers can use to improve
behaviour? These must be the focus of the policy design effort, but
the levers that have usually been used may not be those which produce
the best behaviour, so the analyst may be looking for policy redesign
opportunities.
These points will be briefly mentioned in each case study, but the reader
should bear them continually in mind.
THE CONSULTING FIRM'S PROBLEM
In Fig. 2.11 we studied the flows of trainees and qualified consultants in a

very simple version of a management consultancy firm. We shall now de-
velop a more realistic portrayal of such a firm, starting with a narrative
description.
Narrative description
The management of a consultancy firm has been concerned about the
firm's seeming inability over the past few years to grow with the market
for consultancy services. The firm also seems never to have the right
numbers of qualified consultants and trainees to offer a balanced ser-
vice on existing jobs and to obtain new business.
The firm recruits new staff who require two years of training before
they are considered to be qualified. At the end of the training, a
proportion of the trainees leave, for various reasons, the rest becoming
full-fledged consultants . Qualified consultants stay with the firm for
about 10 years on average before they leave to pursue other careers.
For every 10 trainees, a qualified consultant acts as tutor, and tutors
and trainees are able to work on projects, though they are only about
half as productive as qualified people who are not acting as tutors.
Qualified consultants can move freely between acting as tutors or
working on projects.
Projects tend to be similar in size and, on average, take about 5
months to complete once work has started, but the actual time depends
on the effort the firm can deploy. On average, each project requires a
50 Influence diagram case studies
certain number of qualified consultants, or their equivalent in trainees

and tutors, but, if the actual number of equivalent consultants is less
than the number required for the projects being undertaken, jobs tend
to take longer than normal , and vice versa.
There is an industry standard practice that each project requires a
certa in number of qualified consultants who are not also acting as
tutors. If the number of projects the firm is undertaking at a given time
is higher than this standard, the firm tends to be less successful at
obtaining new projects, and conversely. Otherwise , the rate of obtain-
ing new projects is governed by an exogenous demand for consultancy
services which grows slowly and tends to rise and fall roughly in line
with the business cycle.
Once a project has been obtained, there is a delay of, on average , two
months before work is started, due to the preparation of contracts etc.
The number of projects currently being worked on governs the
number of qualified consultants and equivalents the firm should have.
The discrepancy between that and the actual number is a component of
the desired recruitment rate. The firm wishes, however, to keep up a
reasonably steady stream of recruitment so that the average current
recruitment rate is also a component of the desired recruitment rate .
The firm also has a policy of not dismissing staff on the grounds of
redundancy, so trainees are taken on if the desired recruitment rate is
positive, but people are not dismissed if the desired rate is negative.
Getting started on the problem

The first step is to parse the narrative, being alert for the same concept
being expressed in different words (a typical phenomenon in practical
work), detecting the entities, and trying to spot common modules. It may be
a good idea to make notes by drawing parts of the diagram. These will help
when it comes to applying list extension, and a good variable for the model
list might be the number of projects currently being undertaken.
It is also necessary to look for the policy areas of the model. In fact, there
is only one, the recruitment of new trainees .
In two parts of the narrative, the consequences in the paradigm produce
ungovernable behaviour, stemming from the characteristics of the market
for consultancy services. The first is the extent to which the number of
qualified consultants who are not acting as tutors and the number of
projects currently on hand act to influence the rate at which the firm obtains
new projects. The second arises from the influence that the total workforce
exerts on how long the average project takes to complete.
Both arise because of a mismatch between required and actual resources
and, as the discrepancy between available and required staff was mentioned
as a driver of the recruitment process, it is easy to fall into the trap of
thinking that a discrepancy is the right concept to apply in these two
instances . In fact, when dealing with behavioural effects in managed sys-
tems, it is usually more effective to think in terms of the ratio of desired and
actual. To see that, think in terms of a desired value of 100 and an actual
value of 99. For the recruitment sector, which controls a physical flow, that
would lead to a slow rate of recruitment which would be exactly the same
rate if the two values were 2 and 1, the discrepancy being the same. For the
behavioural effects, the difference between 99 and 100 is obviously far less
serious than the difference between 1 and 2; this clearly seen when we think
of ratios of 99% and 50% respectively.
A suggested solution to the problem is shown in Fig. 3.1, but readers will
essentially be wasting their time if they look at that without first having had
a try at formulating a diagram. That will be a bit of a struggle, but worth the
effort. It is suggested that readers spend about an hour trying before look-
ing at the solution.
A suggested solution
Notice, first, that Fig. 3.1 contains only one place at which links cross. It is,
indeed, impossible to draw the diagram with no crossing links. It is very
unlikely that the reader's solution is as neat, and time must be taken to
redraw diagrams so that they are clear and contain as few crossing links as
possible. If this is not done, the diagram is likely to confuse rather than
illuminate and a diagram which resembles a plate of spaghetti in a bad
temper is poor system dynamics practice.
As before, solid lines show physical flows and broken lines show informa-
tion, actions and behavioural effects. Parameters are in italics and without
signs.
Figure 3.1 is drawn in terms of the physical flows of three entities
(trainees, qualified staff and projects), although the physical flows of
trainees and consultants are measured in a common unit ; they are said to
have the same dimensions.
Both projects and people have essentially the same structure of two
delayed modules in sequence, though the number of projects obtained but
not yet started is not shown, because it was not mentioned in the narrative.
As was stated in Chapter 2, such omissions from narratives are very com-
mon. In this problem, the number of projects being worked on drives the
recruitment process, but the number obtained but not yet started is ignored
in that decision. That may have a significant effect on the ability of the firm
to grow with the market while surviving the oscillations of the business
cycle, so perhaps the number of projects pending should be included in the
recruitment decision; a good example of the qualitative analysis which was
mentioned in Fig. 1.3. The reader should test this, and other policy options,
when we undertake policy analysis for this problem in Chapter 6.
Most of the delays in the problem are shown by a simple D on the link
with an identifying subscript, as was done earlier. This implies that the
- - - -- - - --- - - - -- --- --- --
, ,
,
.... ....
....
....
Ilasoc
""
DefT'lCl')d fo r
Consuttancy
"\ \
SoNIC6S
\
I
I
I
· - ---.:.. Effo rt _ / ..... ~ ,
Ratlo - -- - - 1'roJ«;t
IlM I
,
I
I
" N«maI
" I
I
-----_ _ - - - R~~
"" <-"J.;::;.....
_cI ---
c~.
".,1'IoJ«:I Quohhed .. _ /
- - - Cons.utta nh. _
per Pro iect
-It'
cI_ ~+
aJd~ I
I /
/
/
I - - -- - -
-_/
\ / - - -- --- -
I
\+ /
R«""",,'
Av.~
,.-
Fig.3.1 The consulting firm's problem.
The consulting firm 's problem 53
delay's duration is constant ; in terms of the water flow analogy, the length
of the pipe is fixed.
The narrative states , however, that the delay between project starts and
project completions is not constant; the pipe is, as it were, elastic and can
change its length. That is shown by the arrow and sign leading to DpROJECf
from Effort Ratio! and by the arrow from Standard Project Time-. These
express the narrative statement that if the Equivalent Consultants per
Project , to allow for the lower productivity of trainees and supervisors , are
not sufficient for the projects being undertaken then projects take longer
than normal. This is really only common sense, and painstaking thought and
common sense goes a long way in system dynamics.
The Effort Ratio derives from the workforce available per project rela-
tive to that required per project, and the Equivalent Consultant Workforce
makes allowance for the lower productivity of trainees and tutors. Notice an
example of the limitations of influence diagrams (and of level/rate dia-
grams). The Equivalent Consultant Workforce is shown as being influenced
by three variables and one parameter. The productivity parameter is neces-
sary because it is equivalent consultants the firm wants, but trainees which
it recruits. However, there is no indication of how these four influences
determine the variable.
It is, however, evident that the relationship must be:
EWF= QCAFP+ (QCST + TCBT) *POTAS
where EWF means Equivalent Consultant Workforce , QCAFP means
Qualified Consultant Available for Projects, QCST means Qualified Con-
sultants Supervising Trainees, TCBT means Trainee Consultants Being
Trained and POT AS means Productivity of Trainees and Supervisors, but
this formula cannot be included in the diagram without losing the clarity of
its portrayal of the forces at work in a system. The more sophisticated
software packages for system dynamics do allow one to browse through
portions of the influence diagram and the corresponding equations and
definitions.
As with the Effort Ratio , a Consultants per Project Ratio is defined as a
means of modelling the behavioural influence on the Rate of Obtaining
New Projects. Notice that Basic Demand for Consultancy Services drives
the system with a behavioural link, not a physical flow, the reason being that
the physical flow only starts when projects are obtained.
The link from Desired Recruitment Rate to Trainee Recruitment Rate
has a II symbol to show that there is a non-negativity constraint: trainees are
only recruited when the Desired Recruitment Rate is positive and they are
I We shall adopt the convention of using initial capital letters for the names of variables which
appear on influence diagram s.
2 As before, the reader may well have used different name s for variables and parameters, but
that doe s not matter.
not dismissed, or negatively recruited, when there are too many staff for the
projects on hand.
With these comments in mind, the reader should carefully study this
solution and his own version and ensure that he fully understands what has
been done .
The level of the diagram

How does Fig. 3.1 relate to the cone of influence diagrams displayed in Fig.
2.16?
Its level of detail, showing many parameters, is such that it is practically
at Level 4. As we shall see in Chapter 6, it is only necessary to give
numerical values to the parameters, to specify the rate of increase in de-
mand and the severity of its business cycle oscillation and to define in detail
the effects of the two behavioural links to have a complete simulation
model. In short, it is a very detailed depiction of the problem and, as such,
may be too detailed for some purposes . The reader should certainly attempt
to redraw the diagram to show less detail but to captur e essential features
for, say, the senior partner in this firm. To do so, it may help to study Figs.
2.14 and 2.15.
Qualitative analysis of the diagram

Qualitative analysis was mentioned in Fig. 1.3 as an important component
of the system dynamics methodology. In essence, it involves studying the
diagram to see what can be learned about its adequacy as a model and its
qualities as a control system. These are very broad ideas which are hard to
pin down until one has had some practice.
In this instance, there is only one policy area by which the system can be
managed, the recruitment of trainees. As the model stands', recruitment is
governed only by the number of projects in progress and the size of the
workforce , including trainees and tutors . An average project takes 5
months, but there is also a 2 month delay in signing contracts . No account
is taken of the projects in the contract pipeline , so about 2/7 of the firm's
order book is ignored in determining the manning requirements. When
trainees are recruited, qualified people are immediately assigned to super-
vise them, which has an effect on the availability of qualified consultants not
acting as tutors, but they are the people who have an influence on the ability
to generate new projects. This line of thought suggests that there are other
policy options which the firm might use in controlling recruitment. After
studying Chapters 6 and 7, the reader should experiment to find policies
which have a beneficial effect on corporate performance.
3 As with the production planning model in Chapter 2, this is based on a real problem in which
the policies were as described in the narrative.
The feedback structure of the model
Loops are important to the dynamic behaviour of models and Fig. 3.2
highlights some of the loops in Fig. 3.1 by drawing them with different
patterns of line. The reader should follow the loops by eye, noting that
Loop 4 has a common path with Loop 3 between Qualified Consultants
Available for Projects and Qualified Consultants per Project. Similarly,
Loop 3 is common with Loop 2 from Qualified Consultants per Project via
Projects per Consultant Ratio and thence to Number of Projects in Progress
and is common with Loop 1 from manpower Discrepancy to Trainee Con-
sultants Being Trained. In tracing these loops, take care to follow them in the
direction of the arrows.
Loops 1 to 3 are negative, goal-seeking, loops. Loop 1 controls recruit-
ment to make sure the firm has enough staff. Loop 2, however , acts to
reduce business when there are not enough qualified staff available to
generate it, or generates business to absorb those staff. Loop 3 acts in a
similar way to Loop 2. Loop 4, which is positive or growth-producing,
should act to increase the firm's business in that the more qualified consult-
ants there are, the more projects should be obtained, which should support
more staff. The reader should think round the loops, confirming these ideas.
However, there are some snags in this system of loops which cast doubt
on how well they will function.
Although Loop 1 should ensure that recruitment does not go too far it
competes with a loop, not marked in Fig. 3.2, which runs from Trainee
Recruitment Rate, through Average Recruitment Rate and back to the
start. That appears to be a positive loop which would drive recruitment ever
higher, or lower. The unmarked loop exerts an 'inertia' effect which ought
to be counteracted by Loop 1. Notice that the non-negativity constraint,
which operates in Loops 1,3 and 4 means that these act only when there is
a shortage of manpower. However, as soon as trainees are recruited, some
qualified consultants are immediately switched to supervision, reducing the
resources available to Loop 4 to generate business. Further, the long delay ,
D TR A1N , also means that Loop 4 has little chance to get started.
This kind of qualitative thinking about a system helps to generate ideas
for policy design which are tested in Phase 5 of the system dynamics ap-
proach which was discussed in Figs. 1.3 and, especially, 1.4.
Other applications of this model
System dynamics models are so easy to build that it is nearly always very
bad practice to take a model built for one purpose and adapt it to deal with
another problem. It is usually just as quick , and more satisfactory, to build
a new model for the new problem.
Having said that, it is evident that there is a large class of manpower
planning problems to which the ideas used in this model could be applied.
... .. . . . ............ .... ------
- - - - - -...:., . . I " ... ...
//
- (_N
="Of.c Loop 2 .....
" II"l Progress ...... .....
/ / .....
/ /
"\
II Protect + I + Rate o f + BasK:
"
Demond for "\
I CR~tion II D N~~~~;t ... B CONrAACY- NeC:t~g~ts ...-- Consuttoocy
I _. PROJEC T ~
Services \
Y+ , , " A+ I
~~~f + / " I
• Cons<Jto nt _ _ _ .. Equivalent
AC~d + / I
EQ\JNOIen" / ......... ~d
.... I .... ConsuI talb - - - -
_ Pfoject _ ..... Effort
Ratio _ -- ~ ,
..... .......... ,
-
I "
\
-\.....
,
,
- .
I
("'t'V'\•• ,ltnnt'l:
,
+Y " -, -.-
NutN»r 01
,
\ '- ~- - I
,---- ---
' II" -ed COI'IIiCACJnI.
\ ----------- - - -
Loop 1 EquivOlen t + CotlalManl. ".,~ _ C=:::,,~"
CoroUtonl .. - - _ _ "., ~ _Project
_ • WoncfOl'c e
fJ ..
- - _
.....
- .Qualified
- --- .... +
+ +,
/
-~
I
_to
Cotrecl I
=I "\ . ~C=::'''
lor Plojech +
~ =-.• -•••
.... • ... Loop 3
I
..... .. . .... . ..
"""'_Y I
\ <,
, .
I
, + I
Qualified
CoN<Jltonts .. ~_
_ 01
' .....
- . . Desired,! Su~ 10_ "
Recruitment Loop 4 \
/ ..... ~ Rate +,.
/ \
/ hopCf/lon Quali fied
I Co nsuttants
+
___ . . . Avcik:Jble for
I ___ Projecb Of
I Training
Average Trained \ r
Rectuilmenl
Rate
C~E?"~
- ~ ~
\ "
• \+
\ "\+ ,I
\
" ' .....
\ .....
Tro lne e TransitIOn
\ -- -
(
Recruitment _ _ _ Clo- _ _ . . . Rate to
D EMPLOY + II C~~~t
\ rote t::7fRAIN QuoIifted Rat e
Stat us
Ilecru""""
A ••/alI*Ig
-
Fig.3.2 The consulting firm's feedback structure.
The training of airline pilots is an obvious example. Similarly, many large

firms, and especially the civil and uniformed services, have quite complex
career structures in which one moves up through several ranks , promotion
sometimes being more or less automatic in the early stages but becoming
much more competitive at the higher levels.
These problems can quickly become quite complicated . For example,
there is usually a requirement that there must be some ratio of people at
one rank to those at the next lower rank, that new recruits should have
some known chance of being promoted during their careers and the capac-
ity of the training system may be restricted. At any given rank, it is usually
necessary to serve for some time at that level before becoming eligible for
consideration for further promotion. The reader may wish to try drawing an
outline influence diagram for such a problem. The trick would be to avoid
trying to represent every single rank category and to think instead in
broader terms, such as junior, middle and senior grades.
In these systems, the proportion of people moving to the next higher level
is a management control influenced by circumstances, rather than a param-
eter as in Fig. 3.1. That , of course, suggests the idea that, while the consult-
ing firm may well not be prepared to make people redundant, there is no
reason why the proportion staying after training should depend only on the
preferences of those individuals. Maybe management should have a say in
the matter, in which case the proportion would cease to be a parameter and
become a variable, influenced by, perhaps, the manpower discrepancy.
This style of thinking is a very important characteristic of good policy
design, and the reader should practice developing it as more examples
unfold throughout this book .
THE PHARMACEUTICAL COMPANY'S PROBLEM4
In the previous problem we examined a model in which the system was

heavily influenced by an exogenous force, the Basic Demand for Consul-
tancy Services, which it could do little to influence. The best the firm could
hope for would be to expand its market share, the size of the market being
beyond its control. Such a system is said to be driven or open. In this case
study we shall examine a closed or undriven system, that is, one in which the
behaviour arises entirely within the system itself.
A pharmaceutical company both develops new products and launches
them on to the market. Its products are technically good and the
demand for health care is practically limitless, so the company should
4 Again, based on a real problem, but with the product changed.

be experiencing steady growth. However, there seem to be instabilities

in its behaviour and products which have been developed in the re-
search laboratories are quite frequently abandoned before they can be
launched on the market.
The flow of disposable revenue (revenue after direct manufacturing
and selling costs and overheads) into the firm depends only on the
number of products it has on the market at anyone time and on the
disposable revenue the average product generates. The disposable
revenue for the average product depends on its 'age '; the time for
which it has been on the market. Novel products command much
higher revenues than older ones, partly because of patent protection
and partly because, as products get older, another company will have
time to bring out a younger and perhaps better rival. At the end of their
market lives, products are withdrawn from sale.
The company 's research chemists are competent and, given money
to spend on research, they will, after a considerable time lag, develop
products that are viable in the market place. The company must, how-
ever, incur considerable extra expense and endure further substantial
delay before a product which it has decided to introduce can actually
be offered in the market. Products which have been developed but not
introduced can be thought of as having a certain 'shelf life'. If they have
not been placed on the market, they will have to be abandoned because
other companies will have introduced comparable products.
The company is very marketing-orientated and devotes an increased
fraction of its disposable revenue flow to product introductions as the
average age of its products increases. Whatever is left from the dispos-
able revenue flow is the research budget.
Getting started on the problem

As always, the first step is to parse the problem, and detect entities as a
precursor to list extension or the use of common modules . The reader will
probably find it helpful to do this problem twice using different variables for
the model list. The injunction to think physics still applies!
The tricky area in the model is the research and development sector.
Product research starts lead inevitably to completions, but completed prod-
ucts are not inevitably introduced to the market. Instead, they accumulate
into a pool from which they may be drawn by management to be introduced
to the market, or they may be eliminated as they become obsolete.
As with the consulting firm's problem, the suggested solution in Fig. 3.3 has
been carefully drawn, and in this case there are no crossing lines.
The physical flow for products being introduced and at the market is two
delayed modules . In Fig. 3.3 the products in the DRESEAR CH and DINTROD
delays are shown, even though they were not mentioned in the narrative.
Recall the discussion of this point in the consulting problem. There is no law
about whether the contents of delays should, or should not, be shown in an
influence diagram. The issue is whether doing so clarifies or confuses the
issue and, more importantly, being aware that the contents of delays might
offer options for policy design.
The physics for product research is more subtle. There is, indeed , a
standard delayed flow module for the research phase, after which products
accumulate into a pool of products available for the market. There are,
however, two flows from that pool; one a matter of management action, the
other a consequence of actions.
Product Research
~~~- I:) C001l'01'onRole
-" Start+
/1 1+
Rate - RESEARCH
/ .... ----' -,
\
/ I +// ', +
I I PIOduct AtxtndorYnonl \
I I Rol. "'OdUCllA """bIe
I I A \ _)"forMa1<ellrg
!! :I /
/ <:> \-
: i _lie +
I I _ - - - - - - - - _otSblrQ
I I / / - - ~ PlodJct~l'1tT<>duc1lonl+"
I
, ,
I //
,, +/
Products
" '\
: : ;/ lnln~non INmOD "
: I / _\ + \
I I I "--- _ olCOO1l'01I"O \
I I , "'oductln1Todueftor'o \
i
I
~~"
+~ "" +
/ -, . . ,\' \\
I t ..... .......... ..... \ ,
I I _ Ptoducfs at Martcet \ \
\ 5 \ Pr"2 uct
//
,
\ UFE
I\ 1
I
I
\
\
\
\
,-
\
\
\
\
"'Oducl_
/ / Ilo!.
+ I
I
I
1
1
\\ " -, - \ 1I -tl / II II
\ ..... ..... Average Aoe I I
\ - - - ofProaJc1s ... II I
\ I _ .... ..... _ ;" I
\ I --_ ..../ /
'" //
""
. . . ..... ......
..... _- Froctlonor~
~~~-----
1+
" __ - -- , .. ""
In1roduc1Ioro
Fig. 3.3 Influence diagram for a pharmaceutical company.

The influences on Average Product Age are all negative. At this stage,
that is intuitively correct; introducing new products should reduce age, as
should withdrawing old ones. Similarly, the more products there are in the
market, the lower should be the average age.
The rest of the model is straightforward. The older the products, the
greater the proportion of available money spent on introductions. The
greater the proportion and the greater the flow, the larger will be the Rate
of Starting Product Introductions and the lower the Product Research Start
Rate. It is essential that the reader does not simply glance at these explana-
tions, but follows them closely on the diagram.
The level of the diagram

While it is usually not worth debating in detail the level of any particular
diagram, it is useful to be aware of roughly where it lies on the cone of Fig.
2.16. In this case, the diagram is not at level 4.
There are at least two missing parameters: the respective costs of Product
Research and Product Introductions. The reader should have noticed that
while, for example, Disposable Revenue Flow must be measured in $/
month, the Product Research Start Rate must be in Products/Month. A
parameter of $/Product is missing. If the reader did notice this and did
include these parameters in the model, the reader's diagram is better than
the suggested solution, a good example of not regarding these solutions as
definitive.
Qualitative analysis
There is much else that is missing from this diagram, and the purpose of
giving the reader an incomplete solution is to develop a critical habit of
looking at narratives and models . It is implicitly assumed that the research
capacity and the ability to introduce new products are both unlimited and
completely flexible; the research laboratories can be switched on and off at
will. In Fig. 2.13, point 4, the variable (actual) Products in Research Phase
should make one think of a Desirable or Feasible equivalent, to represent
the ideas of keeping the research team reasonably busy or their physical
capacity, respectively . Similar ideas apply to product introductions.

As with the consultancy company 's model, it is worth taking a closer look at
some of the feedback mechanisms in the pharmaceutical company as shown
in differing patterns of lines in Fig. 3.4. Bear in mind that these are not the
only loops the model contains and the reader should search for, and think
about, more.
Loop 1 is clearly positive and capable of producing growth in the size of
the firm. Since there is no exogenous demand, as there was with the consul-
tancy company , there might be no immediate limit to the growth this loop
could produce, bearing in mind the narrative comment that the demand for
health care products is virtually limitless. In Loop 15, Products at Market
produce Disposable Revenue, so more products can be introduced, and so
on. Loop 3 shares a common path with Loop 2 between Rate of Starting
Product Introductions and Average Age of Products, and is also positive.
As Average Product Age Falls, more revenue is generated per product and
~~~~
( t l'locNct_
PrOduct Research D ..... Comp!e1lonRate
Start Role - RESEARCH \
I
~~ +t
I / ...- --- .... .....
/ I t i l '\ t
I
I
1'locNct~
Rote Ploc1Jc1s_.
\
, f~""MaI<~_
II II - \- \ The Missing
I
I
I
SlIoIIIo t
J Link?
: ~~t
i
I
.~ Phase
1'NT1lOD ~", \
\
I _\ t \
I
I
I
I
I '--=~
I
D1!posoble
Revenue Loop 1 / \ " ' -, \
Flow ~ ---r· \ \ \
ofi' ( --- Of
I
I
I
t~ ..-.......... t \ \ ,
,~~ ') :
\
\
\
\
\
I'
\
\
\ A- \\ \ "./
1'locNct_-t I
\
\ \ ,,- I I
, I
\
\ - ', _~ / I
\
\ • -~~
otProc1Jc1S ~.... / / I
Company
\
\
I -" ",I /
Polley on \ I -------'" '" /
ControlUno
"" ,
~tAge
\ " -, t+ Loop2
"'-, ...... .. rroc:;~~ .-'- - .-

- Spent onProOJct -
- - - - - - - . Infroduc1Ions
Fig. 3.4 Feedback loops for the pharmaceutical company.
5The order in which the loops are numbered is deliberatel y arbitrary and does not imply that
Loop 1 is more important then the other loops mentioned here.
hence more can be spent on introducing new products, driving product age
further down. On the other hand, Loop 2, which is negative, acts to inhibit
Loop 3 in that, whenever products are relatively young, the firm cuts back
on product introductions to leave more money for product research. Fur-
ther , the link from Product Withdrawal Rate to Products at Market con-
nects Loop 2 with Loop 1 and may inhibit Loop 1 from generating its full
growth power.
It is clear that reasoning about loops in this fashion rapidly becomes
cumbersome and it is best seen as a source of insight into aspects of the
model which might repay closer study. In this case, it suggests that the
corporate policy which varies Fraction of Disposable Revenue Flow Spent
on Product Introductions, which was implied in Fig. 3.3 but is explicitly
shown in Fig. 3.4, will need to be worked out very carefully if smooth
growth is to be achieved.
Similarly, following the links up the left-hand side of the diagram leads to
an impasse in that there are no links coming back down and hence there is
no feedback to harmonize the activities of research and marketing. This is
emphasized by the link added from Products Available for Marketing to
Rate of Starting Product Introductions, a link which was missing from the
original system but the addition of which might have a major effect on the
corporate behaviour. We shall test some of these possibilities in a quantita-
tive model in Appendix B.
A PREDATOR/PREY SYSTEM
The dynamics of the relationship between two populations of animals, one

of which preys upon the other, is of great interest in ecological thinking and
has been a fruitful area for analysis. One topic has been the academic need
to understand why populations of animals undergo extreme fluctuations,
another has been to suggest policies by which populations might be man-
aged for economic reasons or to preserve threatened species. In this case
study we shall model the mutual influences of populations of rabbits and
foxes, extending the discussion later to other animal species. We commence
with a very simple narrative. Later in the chapter, the reader will be invited
to improve the model.
A fairly large area of land is inhabited by populations of foxes and

rabbits, but the sizes of the two populations fluctuate quite severely
over a period of years.
The rabbits eat only grass and the rate at which they breed depends
on the size of the population and the inherent fecundity of the species.
Rabbits are killed by foxes and rabbit deaths from natural causes are
very rare.
Foxes eat only rabbits. The rate at which rabbits are killed depends
on the numbers of both species and the efficiency with which foxes
hunt. The rate at which foxes are conceived depends on the numbers of
foxes and their natural fecundity. Since the foxes have no natural
enemies, their life span is determined by the inherent life span of the
species, but life span is reduced if the average food intake of foxes
(from killed rabbits) is less than the food a fox requires for a normal
life span.
Comment on the narrative

This narrative is fairly typical of what might be gleaned from an elementary
textbook on animal biology; that is, various parameters are implied but not
stated and some broad statements are made such as that the rate of rabbit
kills 'depends on' three factors. Part of the skill of influence diagramming
lies in using painstaking common sense to work out whether those depend-
encies are positive or negative .
The narrative also raises questions such as whether it is necessary to
distinguish between male and female foxes, whether one should allow for
immature foxes which are incapable of hunting but which require to be fed
by their parents, as also might pregnant vixens by their partners.
Much else is omitted or implied. For example, it is implied that there is a
limitless supply of grass. The statement that rabbit deaths from natural
causes are very rare excludes both extreme weather conditions and diseases
such as myxomatosis, which is highly infectious and almost inevitably fatal
to rabbits, with obvious effects on fox food supply.
In all such cases, the best advice is to ignore as much detail as possible in
the first attempt at system description via an influence diagram. The model
can always be refined as understanding develops. Refining a model is not
the same thing as making it more complicated, and a model which starts by
being complicated is unlikely ever to become sophisticated.
A suggested solutiou
An initial and, as we have just pointed out , very simple diagram for this
problem appears in Fig. 3.5. This differs from previous diagrams in that
abbreviations for the variable names are given under the text for each. This
is often done as the basis for the connection between the influence diagram
and a subsequent simulation model. The reader should decide whether it
adds to or detracts from the diagram.
The rabbit population sector consists of a simple flow module, with rabbit
births depending on Rabbit Conception Rate delayed by a gestation period.
The physics of fox birth and death are a little more complex as we need to
". RobbII
RGEST
Normd
_
( - -- -
+ ~~4--- -- -~
'r:"
RobbII RABeR
- -PopuIa~ / ~+
~1-
' . . . . ......+_ ..Rabblt ,I
~ ~ -----_ I
~y---------~~R - __ i
FHUF t+ -, I
, ~\r :
J ..... _ - - - - _ AV8I'oge
Normd
R~le
,
~ J »< - ..... "
FecundIY
FOXFfC
,
/
/ "
" AKR
I
I
I( ~ J
I
I
I -:
Po~xpop~on~-
FOX A~
Intake . -+- ..... ./
I / + ', per Fox
AVIPf
:+
I /
:
, I
=
f \
Fox
Concep~ \ I
Rate
FeR \
, <, _ .. ='
Fox
Food
~+
+-___
Normd
\~
+ FOR Food
'\- FSF ~
\ J NF1IPF
~~
\ /
A~+ .... / /
Fox 4-
Rate Ufettme
FaR AFUFE ...
" Normd
" ~
~
NFl1F
Fig. 3.5 A predator/prey system.
allow for the fact that faxes only die from old age, accelerated, perhaps by
shortage of food.
The Fox Death Rate problem introduces a new type of delay. Normally,
an inflow leads to an outflow with a delay between the two. The delay might
be constant or variable but, if the variation arises from the contents of the
delay , we need to show the influences involved explicitly. If we do not, we
shall get the wrong feedback structure. Thus, Fox Death Rate is shown as
depending on the Fox Population and the Average Fox Lifetime, with the
latter deriving from the Fox Population mediated through the food intake
effects. The effects of food intake are that , the greater the Average Food
Intake per Fox, the greater the Food Supply Factor , hence the longer the
Average Fox Lifetime and the lower the Fox Death Rate.
At this point, readers should pause and check that they understand the
reasons for solid lines and dotted lines in the population segments of the
diagram.
An invaluable skill in modelling is to think in terms of the dimensions of
the variables and parameters. In this case, Rabbit Population must be
measured in [RABBITS], where [ ] enclose the units in which the variable
is measured. Similarly, Rabbit Conception Rate must be measured in
[RABBITSNEAR]. Normal Rabbit Fecundity must, therefore, have di-
mensions of [(RABBITSNEAR)/RABBIT], in which we use ( ) within the
[ ] to draw attention to that which flows in the system.
Using a similar line of thought , the Fox Hunting Efficiency must have
dimensions of [l/(FOXES*YEAR)], in other words, the number of rabbits
killed by a fox in a year of normal hunting. The reason is that the narrative
implies that
RABKR=FOXES*RABBITS*FHEFF
Since RABKR must be [RABBITS/YEAR], and foxes and rabbits have
dimensions of [FOXES] and [RABBITS] , ordinary algebra gives the di-
mensions for FHEFF when the variable names in the equation above are
replaced by the dimensions of the known variables and the resulting dimen-
sional equation is solved for the unknown dimensions of FHEFF. As we
shall see in later chapters, dimensional validity is an essential aspect of the
correctness of a model and it is as well to get into the habit of thinking
dimensionally at an early stage.
Average Kill Rate is needed because it is very unlikely that foxes start to
die early simply because today's hunt was not very successful. There must
be a period after which hunger starts to have an effect. Note, however, that
the Averaging Period is a parameter of fox physiology and not, as was the
case with Averaging Period for Order Rate in Fig. 2.7, a matter of manage-
ment choice.
While simple diagrams are generally preferable, this one is too simple for
comfort. Later in this chapter the read er will be asked to extend it to take
account of the food available to rabbits.

This model contains six loops, as shown in Fig. 3.6. Some of the loops have
links in common with others , so the reader should be careful in following
them. Loops Rl and Fl are positive and are capable of producing unlimited
growth in the two populations in the absence of any limiting factors. Loop
R2 provides some limitation in that , the more rabbits there are, the more
will be killed by foxes. Loop F2 increases fox deaths as fox population
increases, thus slowing down the power of Fl. Loops F3 and F4 are negative
and positive, respectively and they function in rather a complex way. F3's
effect is that, the more foxes there are, the less food there will be for each
and hence the shorter their lives, thus decreasing the population. This is
(
i:~
LoopRI O~.) NotmaI
ee:.,cbbll _ _ _ I1aIX>II
4---
.-
+ R~tion ~
Rabbit RABeR
_ Population 4 +
RABPOP ~~
I
I LoopR2 I A:",
,
FA7lME
\ + I , J
........ Ra~_ I
Fox _ Rate I
I
~y __ - - - - - -
Rim ++
RABKR LoopF4
• I
,
I
".---.. +.. ' AV8I"oge

NotmaI
Fox ;' ..... Kill
Rate
, • I LoopF3 "- AKR

~1Iy
FOXFEC
• I
"-
-".
I
!:Fr
I
Fax _ A~ +
P~~ " Intake ......
per Fax
+ \ \ AVIPf
Loop I
=..
\ F2 I
Conception
Rate
FeR LoopFI \. ~at.
~
Food
T+ NotmQ/
~=
FOR Food
+ ~- FSF ~'
\ I NF1lI¥
\ +
A~~e"", ;'
Rate Ufet1me
AFUFE __
FBR
-, NotmaI
, Fox
IJfeIIme
NFI11
Fig. 3.6 Loops in the predator/prey system.
counteracted by F4 which ensures that more foxes will kill more rabbits
producing food to keep foxes alive for longer to kill yet more rabbits.
Obviously, this conflicts with loop R2, which ensures that the more rabbits
are killed the fewer there will be to kill.
Reasoning about loops in a system as complex as this is clearly of limited
value and we shall have to rely on simulation, in Chapter 6, to understand
it fully. The seeds of complex behaviour are, however, clearly present.

Problems similar to this arise in cases such as fishing policy. How many
fish can safely be harvested and what the consequences are for both fish
and people of fishing quotas versus fishing industry subsidies are obvious
policy areas . A complication arises if the fish have other predators, such as
sharks.
A SIMPLE COMBAT MODEL
System dynamics has proved to be rather good for representing the dynam-
ics of combat. This case study indicates a line of approach to the develop-
ment of an extremely simple combat model which will be considerably
extended later in the chapter.
Two nations, traditionally referred to as Blue, for one's own side, and
Red, for the potential enemy, are hostile, and it is possible that Red
may attack Blue. Blue wishes to improve its combat power with the
aim of deterring Red, but has a limited defence budget. Blue' s problem
is to detect the 'pressure points ' in the system at which defence ex-
penditure might have the greatest effect, or defence cuts the least
effect, on combat power. Blue approaches this problem by modelling
a potential combat , with the aim of assessing the broad effects
of defence improvements", not of predicting the exact outcomes of
battle.
The scenario is that both Blue and Red have forces on the frontier
and others in reserve , some distance away. For both sides, if reserves
are ordered into battle , there is a delay before they can arrive and,
because of limitations in the road network, there is a limit to the rate at
which reserves can be despatched.
Each side inflicts combat losses on the other depending on the rate
at which each man fires, the number of men in action and the propor-
tion of shots which hit a target.
The Red commander wishes to achieve a speedy victory and there-
fore commits his reserves as fast as possible. The Blue commander
hopes to wait for Allied forces to arrive from a neighbouring country
and then to achieve an overwhelming victory. He therefore commits
reserves only when he needs to. His criterion of need is that he does not
wish to be outnumbered by more than 2 to 1 at the point of combat?
Parsing the narrative and consideration of Fig. 2.12 should show that the
physics for each side consists of a simple Type A module, the inflow to
6 Indicating the broad effects of system changes is exactly what system dynamics is good at in
all manner of probl ems.
7 This is a very simplified version of the Battle of Waterlo o.
which is the outflow of a Type B delayed flow module. This is shown by the
solid lines in Fig. 3.7.
The Red commander's policy on reserve commitments is to do so as fast
as possible, so the transport limitation is the only true influence shown on
Reserve Commitment Rate. The link from Red Reserves Remaining is
marked with the two lines for a non-negativity constraint to show that
Red will have to cease committing reserves when there are none left to
commit.
The transport capacity and non-negativity constraints also apply to Blue
with the addition of the need to commit reserves when Blue is outnumbered
on the battlefield by more than an acceptable ratio . This is expressed by
comparing the actual and acceptable ratios and triggering the movement of
reserves when the actual ratio exceeds the desired ratio. The link from
Force Ratio Trigger is also marked with two lines to show that there is a
non-negativity constraint, as was the case with the link from Desired Re-
cruitment Rate to Trainee Recruitment Rate in Fig. 3.1. The two non-
negativity constraints mentioned in this paragraph represent slightly
different things. In the first case, Reserve Commitment Rate stops com-
pletely when there are no more reserves left to commit. In the second,
commitment takes place spasmodically as and when the need arises. Obvi-
ously, the first constraint will take precedence over the second.
The loss rates for the two sides are driven by the total firing rate, meas-
ured in [shots/hour], which is the number of men multiplied by the rate at
which each man fires [(shots/man)/hour], and the casualties inflicted by the
average shot [men/shot]. As mentioned earlier , the sooner the reader devel-
ops the habit of thinking in terms of the dimensions of the variables, the
better. In this case, it is a common error to regard a parameter such as Red/
Blue Shot as a kill probability: a probability has no dimensions, which
would lead to a dimensionally invalid equation.
Pressure points in the system
The purpose of the analysis is to guide Blue's defence planning by identify-

ing the pressure points where expenditure might be beneficial. These points
are shown in ellipses in Fig. 3.8. In particular, Blue might increase its
combat power by spending money on any of the following measures.
• Increasing its own transport capacity by building roads.

• Decreasing its own movement delay by better intelligence on enemy
plans so that the reserves can be positioned more appropriately.
• Reducing Red's transport capacity and increasing his movement
delay by using sabotage teams, building obstacles or buying long range
artillery.
• Providing Blue forces with rapid-fire weapons.
• Protective armour to reduce Blue/Red Shot.
- - - - ~- -
Ac:c~_
-~'.
l1olio -- .... ....
+ ....
\ Forc e Ratio ....
\ _ - - (Rod /Blue) " -,
\ t- \
\ \
-- 1 _ 11lwI lledllGl. \
\\ ,,/ -- I $hoi _ _ - ",FIt. \
\ / \ I " f'w Abl I
\ +,( \ I: J
FotcoRaho \ I Total Red + I
I Trigge r \ I / '""Filing Ro to .. -.... /
I \
\ I I '\ I ,,- - H- - ....
I
I \ I I \ I / "
I \ I I \ ' I \
I \ ~I+ 1 / + oJ _ \
I \
\ ~ Loss c~ .:--- ~~;: .. -f}R---C~~jJ:n~';'Rale • R~~~Q5
I Rote . ~
\
-I..
..,.. \ I
I /
,,
, /
I /
I - - --ft-- .... /
I /
/
"\ ""-,
I I \ "~
I I ~ ""
I BkJe Resef'\lQS _ &ue Re$erve r"'\. ~ Blue Re!l9f\19 + "Blue ... Red l oss
RerTlOrIirlg +--- C OmrT"lltmont Rate ~ ArrivoJ Rote .. Combat Rate
I
\ + '" t I ••
\ " I I +I \
- " I \ I \
" /
...... /
---- - \\ /I "\
-~ \ / \
' ~Total Blue .: « \
Rring Role _tIAIe
• $hoi
11ue_. -'
011. . ,., -
Abl
Fig. 3.7 A simple comba t model.

.-------------
(,us Command.,'s
AccflplobJ. FOfr;B --- -- -
Ralio -,
+ -,
Fore e Raho . - -,
\ ..- - - - (Rad/BJu tl) -,
\ \
\ • t- \
\ " \ R.dRaI. \
\ /" ofFlr. \
\ /
"" \ ",.. PerMan I
\, +,( , \\
: I
Force Ratio
" \ Total Roo + I
Trig g s r \
I I -Firing Rate "" - ....... I
I \ I " I
Information \ ",--H--,
I Collectton I \ I
I \ I I \ I / "-
and Command
I andean/rot \ I \ I I \
I \ • ~ + II ~ - \
\ Blue to ss Rod In + Red Reserve Red Reserv e .. Re d Reserves
I Rate Com bat . . - - Arrival Rate
\ Commitment Ro te Rem olning
-1. -, ~ +
l- \ I
I \ I
I \
\ I
I I
\
I -,
I -,
I -,
-,
I
I ~ "
~~~~e~~~e • C~~~t
\ @:
\ ~ ,JI '" I
\ '" \
-, ",," \
...... - -J _ _ -
\
\
\
, \. ::. Total Blue
FIrin g Ra te
;'
,,
/
Initiol fnilial
R.urve Frontie r
Fo>ee Fotee
~
, ,t
, r
. - -l~~~ - --
Fig. 3.8 Pressure points in the simple combat model.

• More training to improve the marksmanship of Blue soldiers.

• Having a larger total force and/or dividing the total force between the
frontier force and the reserves in different ways.
• Better information collection and command and control to monitor Red
in Combat more closely. At present, Blue waits until he is outnumbered
by more than 2 to 1 before starting to commit reserves. Predicting Red 's
movements in advance might enable better results to be achieved .
All the foregoing involve spend ing mone y on assets. There may, however,
be great benefit from changing the Blue commander's policy on reserve
commitment, as shown by the ellipse round Blue Commander's Acceptable
Force Ratio. In all system dynamics , not just military models, the key is to
find the policy which makes the best use of increased assets, not just to find
the best mix of assets for a preordained policy. As we shall see in later
chapters, optimization provides a powerful means of dealing with this diffi-
cult problem.
The idea of pressure points is a useful way of analysing models, and the
method of this book is to bring in new ideas in a steady stream. The reader
should , therefore, look back at some of the earlier influence diagrams and
check them for pressure points. The predator/prey case might appear to
have none , but that would be very rare, as all systems are to some extent
managed. In that example, people might not wish to see rabbits wiped out
and might shoot foxes to pre vent that occurring. On the other hand , rabbits
can be immensely destructive and reducing excessive populations by releas-
ing into the wild foxes bred in captivity might be a more satisfying policy
than gassing rabbits or spreading diseases among them.
This model has an interesting mixture of loops, as shown in Fig. 3.9.

Loop 1 is positive and produces a runaway victory for whichever side can
gain the upper hand. For example, an increase in Red in Combat produces
a greater firing rate, more Blue losses, fewer Red losses and hence, in effect,
produces a snowball effect in Red 's favour (or in Blue 's, depending on the
situation).
The negative Loop 2 represent Blue 's efforts to prevent this happening
by committing reserves. Loop 3, which has a common path with Loop 2
from Force Ratio to Blue in Combat and with Loop 1 from Blue in Combat
to Red in Combat, is also negative . It acts to supplement Loop 1 in that the
additional Blue troops committed by Loop 1 inflict increased losses on Red ,
thereby helping to bring the Force Ratio under control.
Two loops are marked as false loops. This means that they do not exert
a continuous effect, as does Loop 1, or even the spasmodic effect of Loops
2 and 3. Instead, the y simply reflect that a flow must stop when its source
reservoir is empty.
-~,..
A~FoIce
//alia
---- ..... ....
\ Force Ralio~
\ -
- (Red/Blue)
. -
"" , \ A Side 3
LoopEffect
\ ...- +~
\ ./ _//ale
\
\ ./
- f- -1- S/loI - IWIIan
__ "'_ ,
\. +~ I I" I
FaceRaho
Trigger \ : TOId.~Red~+ I
/ \ £ 1 Rnng Rate I _--+1-"
1 /
\ A False \
Loop 2 1
/ The Regulator I / Loop _ \
\ - ,. + .-.
/ ~ .:...- ~~65:ri: +---9R -~~~~ ..
Rate _ I
~ \ ~ l:" "~r" /
t
I I
-;;- .... 11x -'
/ AFalse '\ Snowball _ toss o::::r;-
I
',--
, ,.1 ,
I Loop '"
B&ue ~8fVes - BkJe,~868fVe _ + Buein Rate
I Remaining . - - Cornmo1TT--.1 Rate
\ ;/I; -Be ..... \ t

~::.j~e_ -+ Corrba )\,
..,/ I
<, /
/
- "
-o::::::t:'
+ Told. Blue hdI/IIUe
. ing Rate S/loI
'Iue//ale /
l
"'::""' -
Fig. 3.9 Loops in the simple combat model.
The decline of the Mayan civilization 73
THE DECLINE OF THE MAYAN CIVILIZATION
The Mayan civilization flourished in the southern part of Central America

for a few hundred years from about 400AD. It produced magnificent
art and great monuments in its numerous cities and their temples and
palaces. Over a short period around 800AD the civilization collapsed and
the cities were abandoned, though recent research suggests that the Maya
people continued to live as farmers in the area for a few centuries
afterwards".
Simplifyingconsiderably, the view of archaeologists (or, at any rate , some
of them) is that Maya society consisted of a ruling elite and a mass of
commoners",
The prestige of the elite was reflected by the magnificence and profusion
of the national monuments, and, when their prestige was adversely affected,
perhaps by loss of trade with neighbouring societies due, apparently, to
warfare and pressure from the north, or by declining agricultural output
due to population pressure , they reacted by dragooning commoners into
building more monuments. Inevitably , fewer people were available to
produce food and the consequent poverty seems to have produced yet
further monument building, possibly in an effort to placate the gods.
Whether the commoners obligingly starved to death while continuing to
build, though working ever more slowly as hunger took its effect, or
whether they decamped or rebelled is not known, but the Maya society
certainly collapsed.
In this case study we have not presented a formal narrative, but the
read er may still wish to draw an influence diagram. Hosler 's version is
shown, with some amendments, in Fig. 3.10. The Prestige of the Elite is
shown as driving the process of monument building, together with one of
Hosler 's parameters, the Elite 's Averaging Time for Monument Construc-
tion, which represents their 'policy' for reacting to events. The flow modules
for population and monument construction provide the skeleton around
which the rest of the model revolves.
The summary can be taken further when we consider the feedback loops
shown in Fig. 3.11. Ideally, the reduced prestige should have been adjusted
by the Control Loop building more monuments. Unfortunately, an exces-
sive attempt to apply that control so reduced the people available to pro-
duce food that the Collapse Loop came into playas people , in effect,
starved more quickly than they could build, and then attempted to build
more to placate the gods (or so runs the archaeological theory , according to
Hosler).
The point of this case study is not whether the theory or the interpreta-
8 The discussion in this section , and Fig. 2.10, are based on Hosler et al. (1977). One of Hosler's
co-authors was a system dynamicist, D. Run ge.
• For a fascinating account of similar social structures among the Mexica (Aztecs) see Thomas
(1993).
--
//1 \"
-,
// ', '/ / -,
,,+ \
I I --- "
II \ "/ / - <, , \
I \ II
I
/ Commoner
Death Rote
-com~commOf'le(
1/ \ + BIrth Rote
\
,
I +t ~ + /? \ \
( J , - - - - - -/ / \ '\
I I I
I I I \ E~~""MJf1InlI \
I, II \ rme~ ,
: I I \ I Monument + I
J I \ I ~ +- - ...... -, I
! ,/ \\ _ / t\ )/ i
: I
I
: / '*'Averoge+
pe~~~....._ - - - - ';;numentr- /
I
:
I
i
:: +\ Food -
/
I
1+ / /
I \ Production....... < , / +... ,;'
I I _ - - - - ... C~rd'n ..-- . . .
\ \ ! \ -/ // ,:
J <, , I
\ \ Froctlonof / I
\ \ CommoneB I
\ \ ., Monument I
\ \ CoNtnJeflon I
\\ \\ ++
\
~
Pr;':'~:f I I
\ \ \ I
\ \ ' I I
\ \ I I I
\
"
\
\
',
\
\
\
per
I
ner
co::c
Desired Monuments; /
~tbn
per Commoner
I
" \ ' , \, " \ I It +

\, \, \ I I
\, \, \, I I
\, ,,\' I I
''''' ....
'...........
"
.........
.........' ,'-.....
.....
-
:: I
Foodpet' _ - - -
.... . : :
//
- - ...commoner-
Fig. 3.10 The decline of the Maya.
tion are correct, but to illustrate the use of influence diagrams and simula-
tion models to provide a framework of understanding another discipline.
The collapse of the Maya is a large topic in American archaeology and
many books have been written on it. This diagram puts onto one piece of
paper a summary of a very complex argument. In their simulation model,
Hosler et al. claim to show that a less severe reaction by the elite to the
decline in their prestige would have averted the collapse.
EXTENDING THE CASE STUDIES
The pharmaceutical company revisited

Having studied Fig. 3.3 the management of the pharmaceutical company
realizes that the narrative description it gave to the consultant was woefully
inadequate. This is an exceedingly common occurrence in practical system
,/ / /
----- -, -, //
/,- -",
-,
1/
/ ~'.. 1 ", - - - ......... \
I \ 1 / " "
/ \ \1 1 '.ct+ \
I -com~rs
C ommoner \
/ ~~~~ II \ + BirthRate \
I +t \ . +. ",/:; \ \
/ I ------ / \ \
I I \ Elite'.AVe«>ging I
I I \ 71mfI1Or_' \
, I \ COnIIruc_ \
I / \ I Monument + I
I I \ I ~~y ..... - -, -, I
I
/ \, - / '*'Averoge + -t\ ) i I I
: ~~=~er~ - - - Monuments-"' '/ I
'
I
:I
"+ \ II Control Loop
t+ / /
\
\
\
\ Pr~O<1~- I _ ..... + \
C~o:,
+ '/ ..- /
\ 1 /' Rot.
\ I -" / t+
\ \ Fraction of I
\ \ Commonerl I
\ \ In Monument I
\ \ Construction I
\ \ I
\
\
\
\
\
\
A Collapse
+ i P''''g::OI I
l~ ,
I
I
\ \ Loop I I
\
\
\
\
\ I
I
I
J
Desired Monuments -;. ./
\ \ per Commoner I
\ \ ConstnJction
\ " per Commoner
\ , t +
\ \ I
\ " I
,
, ,
, /
I
" " /
" ... '" ..... ~ / /
.............. -" /
....... .... ..... ..... od oer _---
..... - .....commoner
Fig. 3.11 Loops in the decline of the Maya.
dynamics and does not mean that the managers or the consultant were
incompetent. It almost invariably comes about because the clarity of a well-
drawn influence diagram helps people to think more clearly about what
they do and enables them to describe it more accurately to someone else; a
very useful property for a diagram to possess.
In this case, the management explains that it does not change the flow of
money into product introductions simply because of the age of the products,
even though the Marketing Director might want to do that. Instead, it
attempts to keep spending in the research labs and in the product launch
department on a reasonably even keel, to avoid wasting money by develop-
ing products which are then abandoned because they have passed their
shelf lives and to keep a reasonably steady pool of Products at Market. The
spending problem is approached on the following lines:
• Management takes an average of the product withdrawal rate and bases
the required rate of spending on product introductions on that, in an
attempt to replace old products with new ones. However, it reduces the
required rate as the number of products on the market increases.
• Similarly, management arrives at an idea of the rate of spending on
product research by smoothing out the rate of starting product introduc-
tions, to try to provide for newly-researched products to replace those
being introduced. However, it takes account of products available for
marketing to try to avoid wasting money.
• The total of these two required spending rates may be more or less than
the disposable revenue, but the two required rates enable management
to calculate the fraction of the revenue flow which ought to be spent on
product introductions.
• Having done these calculations, the company is still very market-orien-
tated, so the average product age is used to sway the ratio in favour of
product introductions as the age rises and falls.
This revised narrative is represented in Fig. 3.12, but readers should attempt
the revision themselves before studying the answer. They should, however,
study Fig. 3.12 rather carefully to see exactly what has been changed and
how those changes conform to the new narrative.
Fig. 3.12 is, perhaps, at level 3 in the cone because are still at least six
missing parameters. These are the respective unit costs of developing and
introducing an average product, which were missing from the earlier dia-
gram, the two parameters for the times over which the averages are taken,
and two more which reflect the time constant over which Products Avail-
able for Marketing and Products at Market will be spread in assessing the
two required spending rates . This is clearly seen by thinking in terms of the
dimensions of the variables. For example, if RRSPI is Required Rate of
Spending on Product Introductions, UCPI means Unit Cost of Product
Introductions, APWR stands for Average Product Withdrawal rate and
PAM symbolizes Products at Market we can write the equation implied by
the diagram and then write the equation again, replacing variable names
with the respective dimensions in [ ]:
RRSPI = UCPI*(APWR- PAM)
[$/month] = [$/Product]*([Products/Month] - [Products])
We can now see that the second equation is not dimensionally consistent
as it involves subtracting [Products] from [Products/Month], which is not
legitimate.
Although these parameters would have to be given in a simulation model,
the purpose of an influence diagram is to promote communication and
thinking, and Fig. 3.12 does that.
For instance, the reader should trace the new loops and see that they
provide the equivalent of the missing link which was identified in Fig. 3.4:
Products Available for Marketing does now influence, indirectly, Rate of
Starting Product Introductions. However, there is still no concept of targets
_-'~~~ch
/'
/' I -
I /
I I
I /
A clUal5pendng + I I
on ProductlnlrOduc""", ~ _ I /
~ + r- ~~~
I
I
I "\ /
I ,, " + beSpenton
I
K~;-
I \ 1/ " 1nIr="'n:"..
I
I
\ "t..- ~ +
\ /' I \
I \ /'/' I \
I
I A/' I \
I I \ I \
I I I \ , \
I I \ I \
I I \ J ~o:\ ot ; _
I / \' ProduCllnlToductIons - - - _
I I \ J ++
I I \ I \
I I \ / \
I I Disposable \
I Revenue ...- _ _ _ _ _ \
I I Flow + - ---_ , _ _
I I +. \ ------- _
\ I \ "
\ I \ "
\ 1 \ Ave<ooe Product
\ \ ' WI1hd''''''''' Ilote /'
\ DIspooabIe + It
\
\
\ Reven.Je \ I
/'
---
\ Flow PElf ....
\ \ Product '- I +
\ \ 4 - \ _ Product Withdrawal
\ - -\ - - - - - - - - /' Ilote
\ \\ " "-
/"
\ \ " -,. /
",
\ \ ' - ¥ ....
\
\ \ -- - _ .... Ave<ogeAae
\ - - - - - - - - - - - - - - - - - - olProducIS
\ I . ,
/" _ <,
"",
-,
/
.... ""
" -, I
.... /'
.... ' ............ ..... I
.... .... + ./
-, - ' , PtElSSlJ'6 to Increase
...... ....... ...... - PrOducllnlToducl1on
~'t:
lnaeoses
-- ------ -- --- -
------ -- -- - - -
Fig. 3.12 An extended influence diagram for the pharmaceutical company probl em.
for Products Available for Marketing or Products at Market, so the nega-

tive feedback loops are still rather ill-designed.
Indeed, the very idea of using Products at Market to reduce the Desired
Rate of Spending on Product Introductions rather conflicts with the compa-
ny's aim of corporate growth, which would require positive feedback , not
negative . This is an example of revisiting the statement of the problem
(Step 1 in Fig. 1.3), when study of the system description reveals apparent
conflicts between what the system is said to be supposed to do and what its
structure shows that it is trying to do. Of course, it may also be the case that
the aim of growth is correct and that the Marketing Department is hinder-
ing that by adopting poor policies. This sort of ambiguity is very typical of
practical problem solving.
The food supply of rabbits

The predator/prey model was rather too simple, so a useful exercise is to
extend it to allow for the grass which the rabbits eat , taking account of the
following facts.
At anyone time there is a certain amount of grass available in the area.
The rat e at which grass grows depends on the weather, and the rate at which
it is eaten depends both on the number of rabbits and the grass available per
rabbit. If there is abundant grass the rabbits gorge themselves, but if there
is little they spend time searching for it and fighting over it.
To achieve its normal life span , a rabbit needs a certain average rate of
food consumption. If it eats less, it will be likely to die earlier, unless, of
course , a fox gets it first.
The suggested solution shown in Fig. 3.13 should present few surprises. A
third flow module has appeared for the new entity of grass. The effect of
food intake on rabbit life span is modelled in very much the same way as for
faxes.
The main difficulty is in correctly relating Rabbit Death Rate and Aver-
age Life Span. In the case of the faxes , we could model Fox Death Rate as
depending only on Fox Birth Rate, though the length of life might vary. In
effect, faxes die only as result of having been born, since they have no
predators. Rabbits, however, die either as a result of having been born or by
being killed by faxes. The flow of water goes in to the pipe at one end , but
there is a leak for those killed by faxes , before the rest comes out at the
other end. For influence diagram purposes the form shown in Fig. 3.13 is
correct and, as we shall see in Chapter 8, it is also consistent with the
techn iques for writing equations for a leaky delay. Such effects are quite
common in managed systems (the rejection of low-quality items during
manufacture is one example), so it is important that the reader studies the
solution and learns to recognize when it should be used.
Perhaps the food intake also has an effect on rabbit fecundity . If such is
_-- -__ GrOS$
+ /, "" Consunpfton
~ ~ ..... ~~~
_
c~ogetion
~e
I · +
\
- Gr~
I I ~
+
,.
~C_
RabbIt " -
', ' GRASS
Avai lable
If
I
_ , I
I , " II
-~ !:!-SOa> ' I <, Food + ,. / Grass
Food ~~ \ I - - _ Avolabkt ... _ GrONth
5uppIy AV_ I __ . per RabbIt Rate
RedJcIlon \ \ I ,. - FAPIl GGR
~ I I I .: ~
FSRFR., ~ 1 \ / \
, Avet 0fiJ8 1\ I \
\ " ~?~'_ . . \\(i

\ -~{+ ' I I'
5 ,+ I ~I
.
\ e---:\ \\\ + RGEST _
i- \
~ RDR _ Rabbit _
ConunI>/IOn / - ~\ RabbIt / / /'r::!:!
1ld.1ot / Conc ep tion ..
~. : " ~R
\ -, +~ Av!!!!!!.'11
""", + '-----~/ ~~
H~ ~--- - -- - ---------_ ~n~
E~---- - -_--. Rote ...., /
FHEFF AABKR -, , I
t+ +' ...
I _ -- ---_ Average
_ ' / Kill
Fen I I' -, , Rate
~ Fox' - , Alf
I ....- -- - - ~~~ ' .. I
: ,/ + \ A~+
.+ ,
/'
:
I I \ Inl ok e ..... - .....
I per Fox
y Fox ' Fox AVlPf
~Ilon _:~=
----s::-
' ...
FC~R+ Fox
F\~\ Avetoge +
~ ~SF
..
_ Fox
Blrth lJf~ __ / NFIIPf
~ AFUF£ ....
-- -- -~ ur._
NFUF
Fig. 3.13 Adding the prey's food supply to the pred ator/pr ey system.
the case, a variable for Actual Rabbit Fecundity would be needed, with
appropriate influences from food intake. Howev er, food intake might not
have the same effect on fecundity as it doe s on life span. As influence
diagram practice, the reader should experiment with altering the diagram.
In practical modelling one always has to be careful about extending a
diagram too much, as data for the parameters and effects might be quite
hard to come by.
In the discussion of pressure points refer ence was made to the possibility
of culling faxes to keep the rabbit population up or releasing them into the
wild to keep the rabbit population down. Add these extra flows to the
diagram , bearing in mind that faxes will now no longer die only as a result
of having been born . Do not forget to allow for a Desired Fox Population,
dep endent, perhaps, on a D esir ed Rabbit Population.
Before leaving this model we shall use it to illustrate a wider problem
in modelling. Adding fox culls based on a comparison of actual and
desired fox populations leads to a diagram which implies that the fox
population is instantaneously and precisely known. That is clearly not the
case, and one should refer to perceived actual population, drawing a link
from actual to perceived population, including a delay to imply a time lag in
perceiving what is going on and adding a new factor to represent distortion
and error in observing populations. Similar ideas would apply to the combat
model: it is very unlikely that the Blue commander knows the actual
strength of his own troops in combat, let alone Red's. The reader should
study how the effect of delay and distortion affects the information parts
of the system dynamics paradigm and how to write equations for such
aspects.
Extending the combat model
The combat model of Fig. 3.8 is very simple and the reader should attempt
to incorporate the following factors:
• Blue's soldiers go into battle carrying negligible amounts of ammunition
and the commander of Blue's support services is responsible for bringing
up fresh supplies. He has a target for the amount of ammunition which
should be available for each man in combat, from which he can calculate
the discrepancy between the total requirement and the total availability.
The reserve ammunition is held some distance from the combat area and
requires the same time to move forward as do troops, and it has to
compete for the same transport capacity, a certain number of rounds
taking up the same space as one man.
• The support services commander and the tactical commander therefore
have to resolve the demands for transport capacity if the one wants
to move ammunition at the same time as the other wishes to commit
reserves.
• The ammunition available per man at the front governs the rate at which
each man fires. If there is plenty of ammunition the individuals fire
rapidly, if there is little they fire more slowly.
• Blue also has long range artillery which cannot be attacked by Red but
which can fire at Red 's supply route . The artillery is not accurate enough
to hit Red troops on the march, but the holes it makes in the road have
the effect of reducing Red's transport capacity and of increasing his
movement delay.
• Red has a certain number of engineer troops who can repair holes in the
road but in doing so, use up engineer supplies.
In parsing this narrative, be careful to distinguish clearly between total

quantities and quantities per man.
Summary 81
A suggested solution is shown in Fig. 3.14. Study it carefully noting that:
• Four new flow modules, for Ammunition, Long Range Ammunition,

Current Amount of Damage (holes in the road), and Engineer Supplies.
Total Blue Rate of Fire obviously depletes the front line ammunition
stock.
• Red Transport Capacity is now a variable.
• Blue Transport Capacity is still a parameter, but now affecting two flows.
• There is no non-negativity constraint from Blue Ammunition Reserve
Remaining to Ammunition Despatch Rate, implying that there is a huge
stock of reserve ammunition. The narrative was silent on that point and
that is, again, typical of real SO problems, not just military ones. One of
the roles of the influence diagram is to make one question the informa-
tion available. If the reserve ammunition supply is so large that it has no
effect, it can be omitted from the problem.
• Similarly, there is a link from Long Range Ammunition Stocks/Gun to
Rate of Fire per Gun. That was not specified in the narrative and illus-
trates the risks of jumping to conclusions , in this case by assuming that
what was said about the behaviour of soldiers in combat is also true of
long range artillery.
• The same concept is used for the Red engineer troops and their
supplies.
• Non-n egativity constraints are not needed from the two ammunition
stocks and the engineer supplies , as the reduction in the individual rate of
fire or use will ensure that fire ceases when there is no ammunition left.
Do not complicate an influence diagram unnecessarily.
• The variable called Percentage of Transport Capacity Allocated to
Troop Movement does not imply that troops always take precedence.
Rephrase that variable to refer to ammunition movement, make changes
to signs as required and note that there is no change to the signs of the
loops which pass through that point. As long as a diagram is carefully
thought out , the types of the loops do not depend on the names chosen
for variables.
• Even though the diagram is now much more complicated, there is only
one point at which links cross. Redrawing a diagram to make it clear is
always worth the effort.
• The diagram is still nearly at level 4, but a few parameters are not shown.
Practice dimensional thinking to work out what they are.
SUMMARY
The illustrations in this chapter are the foundation on which you will build
skill in system description and the knowledge of that skill will add much to
your confidence when building simulation models .
A.,.,..,_ _
"'-Ccm~". --- -- -- .... .... ....
+ -- --- ....
\ Perce Raro ....
\ _ _ - - (Red/Blue) ",,
- \
\ , .... -- - -- t-
, _ /hd hdlla..
,
\
""" dFW I
\ »" \
\ .... I I .... ,.., ..... I
~ +,; I I : J
FOfCe Ratio \ I To.OI Red + ,
/ Trigger \ : /"'""fifingRate ... - ...... , I
/ \
/ \ I " ' / .... - -H - - ....
/
/ I , " / / ' ,
,,
~~+ , / + p \
I
/ _
~~~
"::::::.::::
,
" ~:' c~~ :-- ~~~~: . DR c~=r::ate -. Rr~~es
~ .... ~ , . f ..
I
"'J +
I .... ,
I ....
- , I I
/ / ' I I
I I ' I I
I I .... - -it- - .... " I I
,I I Red Transport Eroneer
, Ca pocJty ,- - ~':K
I' I
/ '' ,,
nonQ
I' I \ .>. t- e';::"... Rema
, I Blue Reserve. _ BkJeReserv e of' Blue Reserve Blue ., Red lO$$ I / --- - -.~~ -
, Remarjng . . . . - - - Commitment Rate
SB • AnivaJ Rate - - - + Combat '
Rote
II , ..... ..... ....... t _- - -:,? , I /I N_dhd
~_ II
,. I .... ./ \ +•
1 •
I "
, I , / \ • + j
I II .... .... / ~Ammun/_ - - .... \ I / \ Rote ot Damage
\ / / / ~._..... -< / I I I ./ / \ Repa ll' per Toom Rat e 0 ' Cons.lXTllOQ
I \ .... I .... / \ I I
\ I .... I .... / I CUllen' A.m ount \ I E~eer $Lppf18S
I I
+\ . , .~ I"" I \ .... \ ofDallage ~ -\ \ /
+ / I
C 'l.otJ:,~ed I ~ / .. I ' .:. ~otalBlue .... \ .,+ . . t·I
to~ Movement I De5wed Amrnl.rition /' fi'ng Rate Rote ot Repoll ing - ,. I
I 1 stee les +, Damage I
_ _ d ~
I\ Bluello'e ~r~
I- II . II - ..I " ofF,re P8f - /
I '" I Blue~ -" M<>'1 ( ~
,
\
,,\
..... _ \ AI'nrTu'lifion Stock" /
+ I \+
\
J --- Rate ot
hnict10g
, DiocrepaflC'L "
"
- r\ - -' ...... .... _ BlueAmmunfhon _ +- Damoge.,
.... ....
" ...... ,\ - .... ..... ~-st-- .. .... .... ....
...... ............ - '-. ---------- _ _ - - - Rote of Ammunitlon ..... ~P*
Blue AmrT'lUl'llton AmmunrtlC)n .. hon "... _ : CoNumpfion.. ~ 8afge SheI

Reserve RefTlOlnng
+--......Despctch Rote Arriv al Rate
.. Anvnun
DB ~ I ,
Rote ot Fire \
per Gun \
~~ng~
-, +
. A~
+. I
,::;.~
,
N...."-rd
Fig. 3.14 Combat model exte nde d to include ammunition supplies and long-range fire effects,
References 83
The main theme of the chapter is that influence diagrams can be applied
to just about any kind of problem in which dynamic behaviour might occur,
whether it be the dynamics of the firm, the ecology, the battlefield or a much
wider socio-economic system. Careful study of the diagrams will reveal a
good deal of commonality between them, even though the topics and the
names of the variables are so different. That commonality is the idea behind
the common modules studied earlier, and the basis of system dynamics is
that managed systems all work in pretty much the same way, that is they
have similar units in their structure, almost regardless of the problem do-
main. Recognizing commonalities, without jumping to conclusions, is an
essential part of the art of influence diagramming and system dynamics.
REFERENCES
Hosler, D., Sabloff, l .A. and Runge , D. (1977) Simulation model development:
a case study of classic Maya collapse, in Social Process in Maya Prehistory (ed
N. Hammond), Academic Press, London .
Thomas , H. (1993) The Conquest of Mexico , Hutchinson , London .
CHAPTER 4
Introduction to simulation
FUNDAMENTAL IDEAS
Having seen how influence diagrams are developed and used for qualitative
analysis, it is time to cross the dotted line which separates Stages 3 and 4 in
Fig. 1.3 and move on to quantitative models .
Influence diagrams are models, in the sense that they represent a real
system by using variable names connected by signed links. A quantitative
model , in contrast, represents a system by using variables in equations. In
practice, there is no hope of 'solving' those equations by normal algebra or
calculus; managed systems are too complicated for that. We therefore make
use of simulation, meaning that we will create a set of equations to repre-
sent the system and then allow the equations to run forward in simulated
time to attempt to mirror the behaviour of the real system as it runs forward
in real time. In principle, one could run the equations forward by pencil and
paper calculation; in practice, a computer and specialized software are used.
The key to quantitative simulat ion is that the equations must do the same
things that the real system would have done, and for the same reasons , if the
model is to be regarded as satisfactory.
In system dynamics, simulation is governed entirely by the passage of
time and is referred to as 'time-step' simulation. The other type of simula-
tion in management science is called 'event-based' and is discussed in the
standard management science textbooks. This chapter introduces the es-
sential ideas of the simulation process used in system dynamics and some of
the basic techniques which we shall need to build models. In a later chapter
we shall show that there is no limit to the complexity which can be repre-
sented in system dynamics, and it is, in fact, possible for a model to be both
time-step and event-based.
The essential idea in time-step simulation is that the model takes a
number of steps along the time axis; as many, in fact, as are required to
simulate reasonably accurately the total period which the modeller wishes
to investigate . Each step is quite short, so that there might be many steps in
total. The step length is a critical factor in good simulation and we shall
discuss in due course how a satisfactory step length is determined. The step
length is always denoted by DT.
At the end of each step, two things happen (automatically in the specialist
system dynamics simulation packages). The first is that the variables which
Fundamental ideas 85
represent the system's state are brought up to date to represent the conse-
quences which have ensued during the previous time step ; special measures
being needed to get the simulation to start the first time step. Next, the
variables which represent the flow of information and the initiation of
action are evaluated and the necessary actions are set in train. What actions,
if any, are to take place will have been stipulated by the modeller when the
equations for the system were written.
To bring this about requires two types of variable. The first represents the
states of the system. In control engineering these are called 'state variables '
while in economics they are termed 'stocks', but system dynamics calls them
'levels'. The second represents the physical flows in the system which arise
as a result of actions and produce the consequences in the information/
action/consequences paradigm. In system dynamics terminology they are
called 'rates'; they have no special name in control theory, but economists
recognize them as 'flows' . These variable types are shown in Fig. 4.1.
Levels and rates are the mathematical basis of system dynamics, but, in
practice , a third type of variable is also used: the auxiliary. These reflect the
detailed steps by which information about the current state, the levels, is
transformed into rates to bring about future change. This is shown in the
inner loop of Fig. 4.1, the right-hand side of which is a dashed line to show
the information/action component of the information/action/consequences
paradigm. The left-hand side of the inner loop is shown by a solid line to
signify physical flow, and it also has an integration sign to show that, the
longer the rate continues to flow and the larger it is, the greater will be the
consequence, as felt by the level. The perfect analogy for this is the quantity
of water in a bath when water flows in or out.
The Past ...- - - - Now - - - - -.....

~ The Future
Whereare
wenow ?- - - -_
level. Stock "

or State <, "\
-, \
\
•
\
How did Auxiliary Where do

we get here ?
I we go next ?
I
/ I
/ I
I
Rate.- / '
'"
'" '"
_ _ _ Forces of.. _ - - '
change
Fig. 4.1 The fundamental variable types.

86 Introduction to simulation
The outer loop in Fig. 4.1 and the text and arrows at the top remind us of
the connection with time. The levels are the current state of the system ,
shown by the word 'Now'. The right-hand side is movement into the future
and the left-hand side is movement from the past. Past changes produced
present conditions; present conditions will produce future change. Rates
are, therefore, the movements across time steps, levels are the thresholds
between time steps.
What of auxiliaries? These are intermediate stages by which the levels
determine the rates and they therefore also exist at the thresholds between
time steps and are 'now ' variables.
It is important to realize that every variable in a system dynamics model
is calculated at every DT . Some of the calculations bring the system up to
date, others move it forward in time.
THE FUNDAMENTAL EQUATIONS OF SYSTEM DYNAMICS
What we have said about variables which exist at points in time and those
which exist across time steps suggests a notation for variables to emphasize
their time dependencies. Figure 4.2 shows a time axis with two of its time
steps drawn to a very exaggerated scale.
The three points are conventionally labelled J, K and L. The step from 1
to K is thus referred to as lK and that from K to Las KL. The time length
between successive points, DT, is very small, perhaps only a fraction of a
week. With these time labels, a level, to which we shall give the name
LEVEL, as though it were an ordinary variable, would be referred to as
LEVEL.l or LEVEL.K, depending on which time point was involved.
System dynamics variable names have, therefore, two parts: the name
itself and the time label. The name for a rate would be RA TE.lK or
RATE.KL, again depending on the time step in question. Since auxiliaries
are 'now' variables , existing at the thresholds, they are labelled
AUXILIARY.K.
Figure 4.2 shows that LEVEL has increased from 1 to K because RATE
flowed into it between land K, DT being so small that RATE can validly be
treated as constant over the time-step. Common sense now leads to the first
of the fundamental equations:
L LEVEL.K = LEVEL.l + DT *RA TE.lK
In words, what is in the level at the present time, K, is what was there at the
previous time, J, plus what flowed in during the interval from 1 to K. The
amount that flowed in was the rate of flow, RATE, multiplied by the time
for which it flowed, DT. To emphasize that this is a level equation, it starts
with the type label L.
This seems very obvious, but it is reinforced when we consider the vital
matter of dimensions, or units of measurement. That can be done by writing
The fundamental equations of system dynamics 87
L~EL.K- <,
AUXI~Ry.K
..,
L RATE.KL
RATE.JK
J JK K KL L
l(
time
DT DT
The Past .... - - - -NOW- - - - .TheFuture
1 L LEVEL.K=LEVEL.J+DT"RATE.JK
The Fundamental
Equations:
1
2 R RATE.KL=f(LEVEL.K,AUXILlARY.K,PARAMETERS)
Fig. 4.2 The fundamental equations of system dynamics.
the equation again with, underneath each term, the appropriate dimen-
sions. In this case, we will use the bath water analogy, LEVEL being not the
depth but the amount of water in the bath. The amount must be measured
in litres, and the rate of flow from the tap (faucet) will be in litres/minute.
DT, being time, must be measured in minutes, though we shall write
'minute' to be consistent with the time unit for RATE. Omitting the type
label and spacing out to show the match between the level equation and its
corresponding dimensional equation, this leads to:
LEVEL.K = LEVEL.l + DT*RATE.lK

litres = litres + minute*litres/minute
By ordinary algebra , the minute for DT cancels out the minute for RATE,
so the dimensional equation becomes
Ii tres = Ii tres + Ii tres
This is called a dimensional identity because the units of measurement are
the same on both sides of the = sign, and it demonstrates that our first
fundamental equation is dimensionally consistent. Before we move to the
second fundamental equation we must explain why dimensional consist-
ency matters.
Those levels in a system which represent aspects such as the amount of
stock, the number of people, the quantity of water or the balance of cash,
are conserved quantities; what the levels contain cannot be lost or created.
Even if water flows out of the bath , it has not been lost, merely transferred
elsewhere, so there must always be a mass balance in the system and hence
in the model of that system. The rates in a model are the forces of change
and must be measured in units which are consistent with the levels they
feed. If we use [ ] to denote the dimensions of a variable, then:
[RATE] = [LEVEL]/[TIME]
and
[LEVEL] = [RATE] *[TIME]
For conservation , a rate must have units of the level it feeds divided by the
units in which TIME is measured. The units of a conserved level must be
those of its inflowing rate multiplied by those of TIME!. That is how a real
physical system works, and a model must work in the same way if it is to do
the same thing as the real system and for the same reasons.
Again, what of auxiliaries? They represent the steps by which a level
causes a rate to flow and must, therefore, have whatever dimensions are
needed to ensure dimensional consistency between rates and levels.
We shall have much more to say about the details of dimensional analysis
when we come to formulate equations for the case models.
The first fundamental equation copes with the consequences of what
happened during the JK time-step. The second is the information/action
phase which will take place at K and produce rates for the impending
step KL. Figure 4.1 emphasized that levels and auxiliaries, together with
any parameters in the model, determine the rates as forces of change, so
the second fundamental equation is much more general in form than the
first:
R RATE.KL=f(LEVEL.K,AUXILIARY.K,PARAMETERS)
where the f implies that there is a function or equation which connects
LEVEL, AUXILIARY and PARAMETERS to RATE, and the type label
is R to emphasize that this is a rate . That equation will model the ways in
which the system makes decisions, given the values of LEVEL and AUX-
ILIARY. (There may, of course, be more than one of each in the equation.)
Notice that in the level equation, the rate has the time label JK on the
right-hand side of the =, because past history is being used to update
present conditions. In the rate equation, the time label is .KL on the left-
hand side of the = to show that future changes are being generated by
present conditions.
The significance of the two fundamental equations is that the RATE
equation can be as complex or as simple as is required by the problem;
provided the modeller can understand what is happening, there are no
constraints on what can be modelled. The LEVEL equations, on the other
hand, ensure that the physical quantities are conserved so that the model
cannot get out of balance. That property of the LEVEL equation contrib-
1 Later in this chapter we shall develop a special type of level equation to represent the
smoothing of information which was shown in Fig. 2.12G. The dimensional rule given above
does not apply to smoothed levels.
Time shift and relabelling 89
utes much to ensuring that the model does the same things as the real
system and for the same reasons , to repeat a slogan which the modeller
should learn by heart and repeat whenever a new equation is written. In
summary, the RATEs are the entrepreneurs driving the system forward and
the LEVELs are the accountants, keeping it in balance.
The LEVEL equation should also be seen as representing the memory of
the system. The size of the cash balance level at time point K is the net effect
of all the cash outflows and inflows which have ever taken place . Later in
this chapter we shall discuss a special form of level, the smoothed level,
which gives a 'memory' for information just as the standard, or 'pure', level
which we have discussed so far is a memory for quantities.
TIME SHIff AND RELABELLING
Figure 4.2 seems to imply that the sequence of time points in a simulation
would have to be labelled N, 0, P and so on, and, even using the Greek and
Hebrew alphabets, one would soon run out of labels. The problem is solved
in the system dynamics languages by realizing that J, K and L are simply
labels for a window in time. As simulated time moves forward, the window
is advanced and can be relabelled as a new J, K and L rather than L, M and
N. This is shown in Fig. 4.3; the current position of the window on the time
axis is the nth step, so the next will be n+ 1, at which point the relabelling
automatically happens.
It is important to understand that this can only happen because there are
two main types of variable, levels and rates. The rates are forces
for change into the future, governed by the current state of affairs, the
levels. As the time windows move forward, what was the future for one time
point becomes the past at the next; in Fig. 4.1, we got where we are now
because of forces for change during the last time step and where we want to
go next determines the forces for change in the impending time step.
Step n
I
J K L
I
X
K
Fig.4.3 Time shift and relabelling in system dynamics simulation.

THE SIMULATION PROCESS
Level equations always come in pairs: the equation itself and an additional
equation called the initial condition. Thus our level equation would be:
L LEVEL.K=LEVEL.J+DT*RATE.JK
as before, together with
N LEVEL=100
The initial condition has the type label N and it has no time label because
this is the value at TIME=O when the simulation starts. In this case the
initial condition , usually called an N equation, states that there are 100 units
in the level to start with.
The simulation process thus starts by loading up all levels with their initial
values and then proceeds as shown in Fig. 4.4.
In step 1, all the equations are sorted into the correct sequence for
calculation. This is something that programmers must do for themselves in
an ordinary language such as BASIC , but it is done automatically by the
1 Sort all equations into the

correct order of calculation
1
2 If TIME =0 then set levels -,
to stipulated Initial values
.--------1~3 Integrate to find levels at TIME K
!
4 Calculate Auxiliariesat K and ...._ _.....
Rates for KL
~
IsTIME =LENGTH ?
6 Produce tables and graphs

from stored values of variables
5 Increase TIME by DT
and relabel all
variables
Fig. 4.4 The system dynamics simulation process.

system dynamics languages, saving much time and effort and allowing the
modeller to concentrate on the model, rather than having also to worry
about the details of calculations. At this stage the model is thoroughly
checked for errors and, if serious ones are found, the process stops. All
being well, step 2 is performed to load the levels with their correct values.
At stage 2 a variable for time is automatically created. This will be in the
form of a level:
L TIME.K=TIME.J + DT
N TIME=O
and provides a clock for the simulation .
Having loaded the levels, attention moves to step 4 to get the rates
moving for the first DT in the simulation. After that , the process loops
round through steps 5, 3 and 4 until TIME has grown to the total
period which the modeller wishes to simulate, which, reasonably enough, is
called LENGTH. Throughout the simulation , values of variables selected
by the modeller will have been saved for output at the end of the simulation .
The languages differ somewhat in the facilities they provide for output.
Such, in essence, is the process, and we shall see much of it later.
A SIMULATION EXAMPLE
Before considering some of the special aspects of simulation, such as delays

and information smoothing , it will reinforce the ideas of this chapter if we
construct a very simple simulation model. This will also serve as another
example of influence diagramming, and the reader might find it helpful to
draw the diagram for the following problem before looking at the solution.
The example is deliberately chosen to be rather unrealistic so that the
reader can concentrate on the modelling lessons and not be distracted by
the details of the problem.
The benevolent ruler of a city obtains food from other places, but he
simply has to take what amounts he can get. As food flows in, it is
placed in a stock from which the ruler issues free supplies to his people.
His policy is to release food at the rate of 10% of the stock; if there are
100 tons in stock, he will release food at a rate of 10 tons/day and so on.
In other words, he tries to spread the available food over the next 10
days. If that is not sufficient for the people, they have to buy their own
food elsewhere.
We shall write a simple model of this problem, assuming that initially
there is no food in stock and that none is arriving. On the tenth day, food
starts to arrive at a rate of 100 tons per day. What happens?
The ruler may be able to arrange for deliveries of an additional 50 tons/
day to start in 120 days time. What would be the effect of that?
Figure 4.5 shows the influence diagram for this simple model. The 'Food
Stock Time' corresponds to the 10 day period for which current stocks are
made to last. It is shown in bold italic because it is a parameter, represent-
ing the ruler's policy.
The first step in modelling is to decide on the types of the variables in the
influence diagram. Are they levels, rates or auxiliaries? This process is
called type assignment. It is always a good idea to start by identifying levels,
because all the rest will follow from that. In this case, the Food in Stock is
a level because it is the accumulation of quantity in the system. The descrip-
tion of the system tells us so. As Fig. 4.2 showed, the variables which fill or
drain levels are rates, so Food Inflow Rate and Food Outflow Rate are both
rates. Each variable needs a name, so we will call them FOOD, FINRA and
FOUTR, respectively. The system dynamics software packages differ in
their conventions for variable names , so readers should consult the user
manual for whichever package they are using.
The equation for food in stock will be
L FOOD.K=FOOD.J+DT*(FINRA.JK-FOUTR.JK)
N FOOD=O
It is essential to notice that, in this equation, FINRA has a positive effect on

FOOD and FOUTR has a negative one . These are exactly the signs on the
corresponding links in Fig. 4.5.
The equation for FOUTR is very simple. The current stock is to last for
10 days, so the outflow at any time is the stock at that time divided by 10:
R FOUTR.KL=FOOD.KIFST
C FST=1O
FST, the food stock time, is the ruler's policy and it is defined as a constant,
constants having the type label C.
Food
stock
T1me
-,
-,
-,
~
Food
+
-... Food
Food
Release
Inflow P In Stock ~ Rate
Rate
I 4+
\
\ /
/
" .......
........... "-----~/
./ /
Fig.4.5 Influence diagram for food problem.

Finally, the food inflow is 0 until TIME=lO, at which point it takes a
sudden step to 100 tons/day. Again, the software packages differ slightly in
the facilities they offer for common dynamic processes, but most of them
include a standard function, STEP , to produce step changes of the type we
require. In this case:
R FINRA.KL=STEP(l00,lO)
would ensure that FINRA would increase from its previous value to 100
when TIME=10. Since there is nothing else in the equation, the previous
value is 0, which will have the required effect. In practice it would be better
to put:
R FINRA.KL=STEP(100,10)+STEP(STEPHT,STEPTM)
C STEPTM=120
C STEPHT=O
to allow for the possibility of the second delivery flow starting at

TIME = 120. At this stage, STEPHT is set to 0, but we can turn it to 50 later
to see the effect of the second inflow.
It was stated earlier that the modeller can write the equations in any
order which is convenient to understand and explain, and the packages will
do the chore of sorting them into the correct order. That is exactly what we
have done here, but you might like to type in the equations in a different
order just to satisfy yourself that the results are not changed.
The complete model is shown in Fig. 4.6 and appears on the disk as FIG4-
7.COS. This model happens to be in the COSMIC conventions, but it would
look very similar in most of the other languages. Some points about Fig. 4.6
should be studied.
• The line numbers , starting from 0, are not part of the model, but are
added by COSMIC when the model is printed out. They have been
retained for ease of explanation.
• Lines 0 and 47 are the start and end of the model. The text in line 0 will
appear as the model title on all output.
• All the model is in lower case type, as it is usually easier to read on the
computer screen.
• The lines starting with 'note' are comments and section headings. It is
good practice to group notes in threes to improve model legibility.
As a rough guide, a well-documented model will devote about 20%
of its lines to explanatory comments. A badly documented model is
useless.
• Lines 22-29 set a value for DT, a topic we discuss later, make the model
run for a LENGTH of 300 days and produce printed tables and plotted
graphs of behaviour. Different languages have slightly different conven-
tions for achieving the corresponding effects. It is essential that all the
o * Figure 4.7 Simple Simulation Model
1 note
2 note file named fig4-7 .cos
3 note
4 note food inflow rate
5 note
6 r finra .kl=step(100,10)+step(stepht,steptm)
7 c stepht=O
8 c steptm=120
9 note
10 note food stocks
11 note
12 I food.k=food.j + dt * (f inr a.jk-f out r.jk)
13 n food=O
14 note
15 note food outflow
16 note
17 r foutr.kl=food.klfst
18 c fst=10
19 note
20 note output and control
21 note
22 c dt=0.25
23 c length=300
24 c pltper= 1
25 c prtper= 10
26 print 1)finra
27 print 2)food
28 print 3)foutr
29 plot food=a(0,2000)/finra=b,foutr=c(0, 150)
30 run steady inflow of food
31 c stepht=50
32 run 50% increase in food inflow
33 note
34 note definitions of variables
35 note
36 d dt=(day) time step in model
37 d finra=(ton/day) food inflow rate
38 d food=(ton) stock of food
39 d foutr=(ton/day) food outflow rate
40 d fst=(day) food stock time
41 d length=(day) simulated duration
42 d pltper=(day) output plotting interval
43 d prtper=(day) output printing interval
44 d stepht=(ton/day) step height in food supply
45 d steptm=(day) time of step increase in food supply
46 d time=(day) simulated time
47 +
No syntax errors detected
==== ======== ===
Fig.4.6 A model for the food problem. ' Figure 4.7' appears in line 0 because this
line is used to generate the caption for Fig. 4.7.
variables in a model be printed out and the values are examined closely.
Simply looking at graphs of a few variables is an excellent way to leave
mistakes in a model.
• Line 30 commands the model to run and line 31 sets STEPHT to 50 to
test the effect of the extra supply, after which another run takes place.
This ability to change the parameters and rerun the model is the basis for
the experimentation phase, Step 5A in Figs. 1.3 and 1.4. All the lan-
guages, including COSMIC, allow this to be done interactively on a PC,
but that would not be conveni ent in a book, so we shall use the traditional
method of presetting the experiments to be performed.
• Lines 36-46 are the definitions of the variables, and COSMIC and one of
the other languages differ considerably from the others in this respect.
The type label for a definition is, obviously, D, which is followed by the
variable name, an = sign and, in ( ), the dimensions of the variable or
constant.
• After the dimensions, the meaning of the variable is defined in words.
Notice that definitions must be given for DT , TIME, LENGTH, and the
plotting and printing controls, PRTPER and PLTPER. The definitions ,
'd statements' as they are called, fulfil two essential purposes. The first is
the documentation of the model so that it can be explained to oneself and
to the sponsor. All system dynamics languages provide this in some form.
The second , which is only available in some of the languages , is automatic
checking of the dimensional consistency of the equations. Dimensional
analysis was briefly mentioned above, but it is essential for a model to be
dimensionally valid, otherwise it may not be doing the same things as the
real system and for the same reasons , and no confidence can be placed in
the output. It is possible to check dimensions by hand calculation, but
software does it far more thoroughly.
• Finally, after line 47 is the encouraging message that no syntax errors
were discovered when the model was checked by the simulation
package.
The results of running the model for the base case and with the extra food
inflow are shown, respectively, in the upper and lower halves of Fig. 4.7.
Again , all the packages produce similar graphs and all should give the same
results for this model. As we shall see many more graphs in the book it is
necessary to spend a moment on understanding the conventions.
• The captions on the graphs are the name of the model , which appeared
in line 0 of Fig. 4.6, and the text comment in the run commands in lines
30 and 32.
• The horizontal axis is TIME, in this case from 0 to 300. The d statement
for TIME is drawn immediately under the axis.
• Three variables were plotted, each of which has its own pattern: a solid
line is used for FOOD, for example. The same pattern is used to draw the
vertical axis and to point below the vertical axis to the name of the
variable . In some places, FINRA and FOUTR have the same values, so
their lines merge .
• In the base case, the upper graph, food starts to arrive at TIME= 10; the
STEP effect is clearly visible. FOOD starts to rise smoothly, as does
FOUTR. The behaviour does not settle down until TIME=80, approxi-
mately 70 days after the food started to arrive . This is called the settling
time of the system and the settling times of managed systems are usually
very long. FOUTR stabilizes at 100 tons/day and FOOD at 1000 tons. It
is clear that this must happen. The system can only be stable when the
inflow and outflow are equal and that can only happen when FOOD is 10
times FOUTR; a consequence of the ruler's policy.
• The mode, or general form of dynamic behaviour, does not change when
the extra food arrives , though the numerical values at stability do. The
settling time is unchanged.
The purpose of this very simple model was to introduce some of the ideas
of simulation. We shall revisit it later to see that it can become very inter-
esting indeed when some more factors are included.
TESTS OF CONFIDENCE IN A MODEL
How do we know that the behaviour in Fig. 4.7 is right? This is an essential
component in building up one's confidence in the model, and there are a
number of steps which apply to all models and which can be illustrated from
this one.
• The influence diagram must correspond to the statement of the
problem.
• The equations must correspond to the influence diagram; in particular
the + or - signs in the equations must match the signs in the influence
diagram.
• The model must be dimensionally valid. This one has been checked by
the software, but readers ought to take each equation and write it out
with the dimensions from the d statements underneath, as was done
earlier with the first fundamental equation, and satisfy themselves that
this is indeed a dimensionally consistent model.
• The model does not produce any ridiculous values, such as a negative
food stock.
• The behaviour of the model is plausible - what it does is what we expect
it to do - and the values at which it stabilizes can be confirmed by simple
arithmetic. These tests are harder to apply to a more complex model , but
they are still necessary .
• The model's masses should balance . This means that the total quantity
of, in this case, food which has entered and left the system, together with
what is still there, should be accounted for. This is achieved by making
Tests of confidence in a model 97
150 2000
11
125~ 1750
1oo-tlII
1500
1250
I: 1000
75-+1
Ii
50+1
750
Ii
25~
500
I: 250
olJ -h-'--l".,.r--b..--+.r.--.q.."..---"""'..---~.--In-r..-'~,----l.,..".----l, 00 .
L
0
II TIME (OAYI SIMULATED TIlE
I! F~O~
--- ----- - F N A
~--- FUR
(TONSI STOCI< r:JF FOOD
(TONS/DAY! FOOD INFl.OIf RATE
(TONS/DAY! FOOD OUTFlOlf RATE
STEADY INFLOW OF FOOD (a)

150 2000
j-/-';-
11
125 I
:, 1750 I;
11 1/
Ii 1500
1/
1oo-tl 1250
r----.7-~--_-l
I: I
:/
I
1000
I: !/
75-+1
50+1
750
:I
II II
25~
500
,,
Ii 250 ,,, I
olJ O-;rL--l:m--\m--im-~..-"'o::r..-~r-\m-r-mr~"'---\'
I' L· TIME IOAY! SI~TEO
I!
TIlE
F90~ (TONS! STOCl< Of' FOOD
....=.=-==~O~T~ m=~g:~l ~~ ~~~R~~
50% INCREASE IN FOOD INFLOW (b)
Fig. 4.7 A simple simulation model.
some additions to the model, which appear in FIG4-7A.COS on the disk.

One adds two new equations for the cumulative inflow and outflow of
food:
L CUMIN.K=CUMIN.J +DT*FINRA.JK
NCUMIN=O
• Even simple equation-writing practice is useful, so readers should write

their own equations for CUMOUT. The initial condition for FOOD is
worth rewriting as:
N FOOD=IFOOD
C IFOOD=O
At first sight, this makes no difference to the model, but it allows the
initial amount of food to be changed in a ~ter run . With these additions,
a check variable can be written:
A MCHECK.K=CUMIN.K-(FOOD.K- IFOOD)-CUMOUT.K
In words, MCHECK compares the total which came in with the change
in the food stock and the total amount which was distributed. Ideally,
MCHECK will be 0; in practice, because of rounding error in the com-
puter, it will be some small number, such as -12.45 X 10- 3 (why does it
not matter that this is a negative number?). Notice in FIG4-7A.COS on
the disk that all the new variables have been added to the printed tables
so that MCHECK can be seen to check, and that they have also been
added to the model documentation so that dimensional analysis can be
done again to make sure that the additions have not undermined the
model 's dimensional consistency.
• We mentioned earlier that the rates in a model can be seen as the
entrepreneurs and the levels as the accountants. On that analogy, mass
balance equation s are the auditors, making sure that the accountants have
done their job properly. These precautions may seem unnecessary for
such a small model , but it is as well to get into good modelling habits from
the start.
Having examined a small model to familiarize ourselves with some of the
procedures, we shall now look at some important components of system
dynamics models.
REPRESENTING DELAYS IN MODELS
We have seen in Chapter 1 that delays are an extremely significant feature

of managed systems, and we examined, in Chapters 2 and 3, how delays can
be included in influence diagrams . We must now develop a method of
representing them in quantitative models .
In the earlier influence diagrams we showed a delay by:
INFLOW - D ---) OUTFLOW
with a + sign on the link. This is simple, but does not draw attention closely
enough to the fact that between INFLOW and OUTFLOW there will be a
quantity of material in the pipeline . If INFLOW is the rate of recruitment
of workers and OUTFLOW is the rate at which they complete their train-
ing, then the 'contents' of the delay are the number of people in the process
of being trained, which will rise and fall as INFLOW and OUTFLOW vary.
CONTENTS is clearly a level variable and is, obviously, fed by INFLOW

and depleted by OUTFLOW.
Figure 4.1, the fundamental variable types, stated that levels control
rates, and, if this is true for dynamic systems as a whole, it must also be true
for delays as important components of such systems. This is shown in Fig.
4.8 for three increasingly complicated cases.
In the first case, the outflow at any time is simply CONTENTS divided by
the magnitude of the delay. Thus, if the delay in training is, on average, 10
days, then the daily rate of completion of training will be 10% of the
number of trainees. This is, of course, also the city ruler's policy for issuing
food; on average there will be a lO-day delay between food arriving and it
being issued.
INFLOW ..
+
CONTENTS .. OUTFLOW
\ + ....
--
\ / \
<, / \
Delay
LCONTENTS.K=CONTENTS.J+DT'(INFLOWJK-OUTFLOWJK)
ROUTFLOWKL=CONTENTS.K/DELAY
A) A First Order Delay
INFLOW----":- CONT1~ TRANSlTION""";CONT2 +-OUTFLOW

\ +r~ \ +t~
<, ./ \ '-./ \
- - DELAY/2
DELAY/2
LCONTl.K=CONT1 .J+DT'(INFLOWJK-TRANSR.JIO
RTRANSR,KL=CONT1 .K/(DELAY12)
LC0NT2.K=CONT2.J+DT·(lRANSR.JK-oUTFLOW,JIO
ROUTFLOWKL=CONT2.K/(DELAY/2)
B) Two FirstOrder Delays 'Cascaded' to Give a Second Order Delay
-- --
+ - + - + -
+;.. \ +A ..
INFLOW-., CONT1"-TRANSR1----.CONT2. . .TRANSR2.....CONT3..... OUTFLOW
\ 1+ \
\ \ -_/ \
DELAY/3 DELAY/3 DELAY/3
L CONT1 .K=CONTl .J+DT'(INFLOWJK-TRANSR1 .JIO
RTRANSR1 .KL=CONT1 .K/(DELAY13)
LCONT2.K=CONT2.J+DT'(TRANSR1 .JK·TRANSR2.JIO
RTRANSR2.KL=CONT2.K/(DELAY13)
L CONT3.K=CONT3.J+DT'(TRANSR2,JK-oUTFLOWJIO
ROUTFLOWKL=CONT3.K/(DELAY/3)
C) Three First Order Delays Give a Third Order Delay
Fig, 4,8 Representing delays in models.

This type of delay is called 'first-order' because only one level is used to
store the contents'. Its behaviour is shown in Fig. 4.93, in which DELAY has
been set to 20 days, in response to a step in INFLOW. OUTI, the response
of the first-order delay is initially immediate, but is eventually sluggish and,
theoretically, never reaches the value of INFLOW; some trainees never
learn .
The first-order delay, for all its simplicity, can be quite a good model for
certain processes. For example , a bus company recruits drivers, all of whom
have the necessary licence to drive a bus and all of whom know the town
fairly well. Some of them will become fully productive quite quickly; others
will take longer to learn their route, the fares and so on.
For other problems, the first-order delay is too simple. The training of
scientists, for example , must take at least three years to complete a first
degree, after which some will be productive quickly, while others will take
longer. Such delays are modelled by 'cascading' as many first-order delays
as required to produce a higher order of delay. Figure 4.8(b) shows a
second-order delay. There are now two internal levels, CONTI and
25 .
...--:.----:==-==----
, e>
: ~~::-----
: /?,,--
12 .
1,,1
/1f
///1
//,.'1 :
/;/ I
/ ;/
!
:
2.
I
/11 .
90 .
III :L
, . . . . o. ~o. ~o . ~o .
TI ME (WEEK) TIME WITHIN SIMULATION
:'II:-- OUT
INFLOW (UNIT/WEEK) INFLOW RATE TO OELAY
I:L=8~H 1 (UNIT/WEEK) OUTPUT FROM FIRST ORDER DELAY
!~~m:~~~l ~~~~ ~=g= ~~~~~Da:~~RD~~~~Y

l -.. _ OUTP (UNIT/WEEK) OUTPUT FROM PIPELINE DELAY
DISPLAY OF DYNAMICS
Fig. 4.9 The dynamics of delays.
2 A first-order delay has the same output shape as an Erlang Type 1 distribution, and corre-
spondingly for the higher order delays discussed below.
3 Notice that the model LENGTH in Fig. 4.7 was 300 days, to allow for the effects of the
second step to be seen. In the rest of this chapter, LENGTH is reduced to 100, which is
sufficient to show the dynamics.
CON1'2, between which there is a transition rate, TRANSR. Notice that
each of the components of CONTENTS controls its successor rate and
that the total delay is split equally between the two stages. This produces
the behaviour shown for OU1'2 in Fig. 4.9. After the step in INFLOW
there is a short dead time before OU1'2 starts to respond, after which the
rise in OU1'2 towards INFLOW is swifter than OUTl's was, though
OU1'2 is only just over half way to INFLOW 's level after the 20 days have
passed.
Cascading three delays gives the 'third-order' response shown for OUT3.
Figure 4.9 suggests that, as the order is increased by cascading more
delays and splitting the total delay across more transition rates, the shape of
the response comes closer to a step delayed by 20 days from INFLOW's
step. This is, indeed , the case and, theoretically, one can create a delay of
infinite order in which OUTFLOW has exactly the shape of INFLOW and
lags behind by exactly the amount of the delay. In system dynamics, an
infinite order delay is called a 'pipeline delay', though some packages use
slightly different terms, and this is shown in Fig. 4.9 by OUTP, which exactly
matches INFLOW exactly 20 days later.
It would obviously be incredibly tedious to write the equations shown in
Fig. 4.9 every time one wished to model a delay, and all the packages
include some standard delay functions. A third-order delay is usually writ-
ten as:
R OUT.KL=DELAY3(IN.JK,DEL)
C DEL=10
The time label .JK within the DELAY3 does not mean that OUT.KL
depends on IN in the preceding time step; it simply reminds one that the
current value of OUT arises from previous values of IN, approximately
DEL time units ago. COSMIC has standard functions for all the delay
types.
The total amount within the delay is modelled as:
L TAMOUNT.K=TAMOUNT.J + DT*(IN.JK-OUT.JK)
but what should be the initial value of TAMOUNT? To answer that we
recall that a DELAY3, for example, has three internal levels. Like all levels
they have to have initial values and these are generated automatically by the
software package. In most cases, the three internal levels of DELAY3 are
each set to an initial value of:
N CONTl=IN*(DEL/3)
and so on. The total amount in the delay will, therefore, be IN*DEL, so the
correct initial condition for TAMOUNT will be:
N TAMOUNT=IN*DEL
otherwise the external level, TAMOUNT, which the user has created, will
not be consistent with the total amount in the three inaccessible levels
within the DELAY3 function. This equation for the initial value of
TAMOUNT would be correct for any order of delay. Initializing delays is a
most important concept to which we shall return when we consider the
modelling of more than one output from a delay.
Although Fig. 4.9 was produced using a step for INFLOW, for simplicity,
the corresponding behaviour patterns are produced for any other shape of
INFLOW. In particular, the output of a pipeline delay will always exactly
match the inflow and will lag behind the inflow by exactly the magnitude of
the delay . The program disk includes FIG4-9.COS, and the reader should
run this program, experimenting with the size of delay and , when we have
studied how to write the equations, with different patterns for INFLOW.
We have shown that different types of delay can be modelled using the
standard functions available in the various software packages, but which
type of delay should one use in any particular case? We shall deal with that
in Chapter 6 when we build models for the case problems.
As we have remarked, delays are a vital component of managed systems
and often have a significant effect on the dynamics, so it is as well to be quite
clear about what is being modelled in a delay. The important point is that
the delays do not represent individual entities, such as particular bus drivers
or bags of food. They do represent the overall flow in the system. This can
lead to apparently strange results in that, while an issue rate of 151.279 tons
of food per day does have meaning , what meaning could be attached to
151.279 drivers completing training per day? At first glance, 0.279 drivers is
nonsense, but the point to grasp is that system dynamics concerns itself,
usually, with the broad behaviour of the system , rather than the fine detail.
When necessary , however, one can model exact integer quantities, as we
shall see later, providing one has thought through whether the extra effort
and detail are really worthwhile , given that the purpose of system dynamics
is strategic policy design for a system.
Before we leave the question of delays, notice that the influence diagram
in Fig. 4.5 is the same as Fig. 4.8(a), the dynamics in Fig. 4.7 are those of
oun in Fig. 4.9, and recall that in Chapter 3 we modelled the elimination
of a discrepancy between desired and actual stocks over a period of time. It
should be clear, therefore, that the ruler's and the production manager's
policies both have the same effect as a first-order delay and that the time
constants in those policies are delays.
SMOOTHING OF INFORMAnON
Suppose that the city ruler decides that making his people wait 10 days for
food is too harsh and instead he would like to issue food at a rate of half the
INFLOW and 10% of the stock. He realizes that, if he issued at the same
rat e as INFLOW there would be no stock to tide people over any irregular-
ities in INFLOW. We are tempted to write an equation:
R FOUTRKL= FOOD.K/FST + FINRA.KL*0.5
but this would constitute a most fundamental error of system dynamics
theory. The reason is that values with the time label .KL are yet to happen
and cannot be used in decisions about what is yet to happen. In common
sense terms , the ruler does not know what food will come in toda y, so he
cannot base today's decision on something unknown. Even if someone rode
ahead of the food convoy yesterday and told him what was expected to
arrive today that would still not be the same as what does arrive today. It
would be information at time K, not flow during KL.
In short, to use a rate variable on the right-hand side of another rate or an
auxiliary is to make the serious error of confusing the future with the
present. Review Fig. 4.1.
The best the ruler could do, therefore, would be to take account of the
average rate at which food had arrived over the past few days, AVFINRA,
and use a policy of:
R FOUTRKL=FOOD.K/FST + AVFINRA.K*0.5
Notice that the time label on AVINFRA is .K because it is something the
ruler knows now.
This process is called smoothing of information and, like delays, it is a
very important component of dynamic systems. It has a standard equation
form which is shown in Fig. 4.1O(a).
The smoothing equation is a level with a definite structure. The differ-
ence between the variable being smoothed and the smoothed variable is
weighted by the effect of the time constant and added to the smoothed
variable. The effect of this is best understood by imagining that FINRA has
been constant for a long time. Clearly, AVINFRA, FINRA's average value,
will be the same as FINRA and will have been the same for a long time.
That corresponds to the system' s 'memory' that FINRA has been stable.
Suppose that FINRA now steps up to a new value and stays there. In the
first few DTs, the memory will not change much, as the system cannot
abandon all its previous history of FINRA and suddenly switch to the new
value. As time passes and FINRA stays at its new value, the system will
come to accept the change as permanent and AVINFRA will move closer
and closer to the new FINRA.
This behaviour is shown in Fig. 4.11, in which FINRA starts at zero and
steps to 100 tons/day , as it did in Fig. 4.7. Now, however, it becomes very
noisy and unpredictable after day 30, because of storms in the mountains.
When the step happens, the smoothed value tracks up to the new value and
then follows behind the variations, more sharply in the upper graph , in
which the smoothing time constant is 4 weeks, than in the lower graph
smoothed variable
/ / <,
/ / ................
/ ........
/ / <,
t/ ~ <, ~
L AVINFRA.K=AVINFRA,J+(DT/TCON)(FINRA.JK-AVINFRA.J)
/' "-
/ "-,,-
/
time constant "-
variable being smoothed
_ _ _ _ variable being smoothed

N AVINFRA-FINRA < Isthe usual/nlfial value
A) The Equation
L AVINFRA.K=AVINFRA.J+(DT/TCON)(FINRA.JK-AVINFRA,J)
(tons/ _ (tons/ + ((day)k d aY))( (tons/ _ (tons/)

day) - day) I' day) day)
B) The Dimensions
Fig. 4.10 The smoothing equation .
where it is 10 weeks. Choosing a suitable time constant is a matter of

balance between accepting a genuine change reasonably quickly while
avoiding chasing the random noise. The model which produces Fig. 4.11 is
included on the model disk, but it employs some modelling techniques
which we shall not study until later. The reader should, however , run the
model with various values of TCON and get a 'feel' for the way in which the
behaviour of AVINFRA changes.
The dimensions of the smoothed level are shown in Fig. 4.10(b). FINRA
has dimensions of [tons/day] and DT and TCON both have dimensions of
days. The effect is that the level, AVINFRA has the same dimensions as the
rate, FINRA. At first sight, this is strange, especially as we stated earlier
that a conserved level must have the dimension of its infl.owing rate multi-
plied by those of time. The reason is that the smoothed level is not a
conserved level. In an equation such as
L TAMOUNT.K=TAMOUNT.J + DT*(IN.JK -OUT.JK)
200
",~
150
125
100
75
50
25
L" " TIME (DAY) SIIU.ATED TINE
L_ ~~~~~RA l~=~~~ ~~=ci~::'RATE

SMOOTHING TIME = 4 (a)
200
175
150
125
100
75
50
25
L" " TIME (DAY) SIIU.ATED TINE
L_~~~~~RA m=~~~l ~~=ci~::'RATE

SMOOTHING TIME = 10 (b)
Fig. 4.11 Smoothing information.
if IN and OUT suddenly become zero, TAMOUNT will not change from its
previous value but will continue to act as a memory for all the flows which
have taken place. In the smoothed equation:
L AVINFRAK =
AVINFRAJ + (DTrrCON)(INFRAJK- AVINFRA.JK)
if INFRA becomes zero , then AVINFRA will eventually decay to zero ,
thus acting as a memory of INFRA's behaviour, while FOOD will act as a
memory of INFRA's consequences. In short, the variable name Average
Food Inflow Rate does not mean that the variable is rate type.
Let us also note that the proposed policy equation:

R FOUTR.KL=FOOD.KlFST+ AVINFRAK*0.5
would be bad modelling practice. The value of 0.5 could not be changed in
a rerun to experiment with different policies and it has no stated dimen-
sions, which may well lead to errors when dimensional consistency is evalu-
ated for the model. It would be far better practice to put:
R FOUTR.KL=FOOD.KIFST+AVINFRAK*FRAC
C FRAC=O.5
together with
D FRAC=(l) fraction of average inflow which will be issued
in which (1) means that FRAC is a dimensionless constant.
POLICY EXPERIMENTS WITH THE FOOD MODEL
Having discussed the equations for smoothing, let us now move into the
policy analysis phase of system dynamics and see what effect this change of
policy has. Model FIG4-12 .COS on the disk is set up to provide this by
adding a 'policy switch' to the equation for FOUTR:
R FOUTR.KL= STOCK.KIFST + PSWITCH*(AVINFRAK*FRAC)
The policy switch is a constant which can be set to 0 or 1. If it is 0 the effect

of AVINFRAK*FRAC is suppressed, whereas , when PSWITCH=l, the
new policy can have its effects. Obviously, one could achieve the same
effect by setting FRAC to 0 or to any other value, but the switch method is
useful in more complicated cases, which is why we mention it now.
The consequences of the change in policy are shown in Fig. 4.12.
Fig. 4.12(a) and (b) show two graphs with the policy off and on. The general
result is that implementing the new policy will lead to generally lower
stocks, which might have been expected. The food outflow rises more
quickly in the first 20 days but, after then, there is a generally less smooth
pattern of food outflow. The new policy seems to have made matters worse
overall.
This can be seen more clearly in Fig. 4.l2(c), which gives comparative
results for the two policies. In these comparative plots the caption states
which variable is being compared and the patterns correspond to the
different runs - the scale line leads to the variable name followed by, not
its definition, but the text on the run statement.
These policy comparisons, even for such a simple model, lead to some
important general comments.
150 2000
11 Ii
:: It
12!1~ 1750
,\ ~ I::
II
n
i\
1\ }, ::
II
II
I: 1500 1\ 'I
j". \ ,' 1
, ~,
II
~
r.
I ' , •
I I
,-- L
I I
1\ lA"
I, I
~
I \ 11 I
: '/! i~'\!.....
,-V"" l '\
rvu
't'-"'"1\,!/
l00-n ---------, I , , " '-J"".~ t ,~ I I
1250
...l-.. 1//r'y...... :T \ II
/: 1000 /
" i 'V
I I
II 1' ,I ', '.! ' .:
"II
7!1-H
I ' ,J '
Ii 750 I
50~ /
Ii 500 /
/,
25-W
I: 250
o II 04.---t.,.,--l.",.-l",--I;".,----l."------t...----!,,.--\,;.,.--j,,,---l, 00.
I: L' TIME (CAY) SUU..ATED TIME.
IL-------.~~2aA
- - - - - FOUTR
m=~o:~~a: ~~QI RATE
(TONS/DAY]'ooo ClJTFLOIf RAT!
OUTFLOW BASED ONLY ON STOCKS OF FOOD (a)
150 2000
11 Ii
I: r:
125~ 1750
,\ ,~ n
11 i\ ::
I, LWt!'
!'!A
I: 1500
p ~" t\. II
10011
::o.Jl/M
:------::;;.::1, '/
I 1\ :
~, '\,;,..
)o!1~)-,..J t,-,\l,' ,....;~ i,
1\ , ' : '
I \
I tI" :,f\:1\
Ii
1250
I I :\V ,
,
I ,"
" , V' ',I
I'
\\,0.,'
r"u",
,
-; II
v
l~ ,I\,
.
75+1 1000 '
J / ' I
, "
I I "II
Ii :, I 't l
. ,/ ~ \~
750
50+1 II
Ii 500 l/
25 -!J :/
II 250
,1/
oll 0 ~o. 30. ~o. 50. Bo. '0 . Bo. 90. ~OO .
I' L-
I!
T I ME (DAY) S IIiLUTED r UE
FOO~ ITONS) STOCK Of' FOOD

...:::===.:~o~9A l~=~:~l ~::: &,~",,"~;.
OUTFLOW BASED ON STOCKS AND AVERAGE INFLOW (b)
2000
17!10
1500
1250
1000
750
500
il~88B
TIME (DAY! .llI.Uml TIlE
OU~ IAIED CN..Y C»I ITOCQ aF "XID
DUTn.. . IAI!D ON ITOCKI IHJ ,,~ INFl.DII
FOOD (TONS) STOCK OF FOOD (c)
Fig. 4.12 Comparison of ruler's policies.

Policy usually has a very marked effect on the behaviour of a system. We

shall see this many times, but it is important for the reader to grasp this now.
It is one reason why it is usually not worthwhile to complicate a model so
that it always deals with integer numbers of, say, bus drivers. Concentrating
on the policy issues is usually a better strategy for bringing about an im-
provement in managed systems.
Systems often exhibit counter-intuitive behaviour, that is, what they do
is often not what one would have expected. It seems obvious that issuing
food as it comes in is going to improve the lot of the people. It turns out
that it does not. Food reserves are smaller and the rate of issue is more
variable.
Policies need not be simple choices. Is there a policy that gets the benefit
of a rapid rise in issue rate as food starts to come in, as the new policy does,
but gives the higher stocks as protection against the future and smoother
behaviour on a daily basis? To test that, the reader should change FIG4-
12.COS as follows:
• Replace FRAC by an auxiliary variable, FRAC.K, in the equation for

FOUTR.KL.
• Replace C FRAC=O.5 by C BFRAC=0.5 to give a base value for the
fraction issued. This will allow BFRAC to be changed in re-runs to allow
one to test alternative values for it.
• Add a new group of equations:
A FRAC.K=BFRAC-STEP(BFRAC,CTIME)
C CTIME=200
As LENGTH is only 100, the STEP will not have any effect, so the
two polices are unchanged. If, however, CTIME is set to, say, 30 in
a run in which PSWITCH=l, one would be testing the new policy
being implemented when the advantage of the more rapid rise
in food issue rate had been enjoyed, after which the policy would
revert to the original, smoother case, or so one would expect. Does
it?
It is essential that you should make these experiments, consulting the

user manual for the package you happen to be using for guidance on exactly
how to do it. You will gain much more from carrying out experiments for
yourself than from any amount of reading about experiments in a book. It
is also essential to document all variables, work out their dimensions and test
the changed model to ensure that it is still dimensionally consistent. This is
such a fundamental of good modelling practice that it should always be
done whenever a model is changed.
Before summarizing this chapter we must consider some aspects of DT,
the step size.
Summary 109
SELECTING A VALUE FOR DT
If DT is too small time will be wasted waiting for models to run. If it is too
large, numerical instability may occur, which means that dynamics in the
model may be due to errors of calculation, not to the dynamics of the
system. A good compromise is to ensure that
DT:o:;;DELI(4*ORDER)
for the smallest delay in the model. Thus, if a model contained a DELAY3
with a delay magnitude of 10 and also contained a time constant of 4 in a
smoothing equation, the candidate values for DTwould be 10/(4X3)=0.833
and 4/(4Xl)=I, because a smoothing process is a first-order delay, as is the
ruler 's policy of FST=lO. However, DT must also be set to one of several
standard values in order to avoid rounding error in the calculation of TIME
affecting the times at which time-dependent actions such as STEP take
place. The standard values are 0.5, 0.25,0.125and so on, so the correct value
in this case would be 0.5. In the model in this chapter we have used 0.25,
which is smaller than strictly necessary, but gives smooth graphs.
It is not wrong to use very small values, such as 0.015625, but it would be
unusual, and such small values should only be used when one has had a
good deal of experience of modelling.
THE SIGNIFICANCE OF DT
The equations in a model are supposed to represent the way in which the
real system works. However, in order to get the equations to run on a
computer, an additional factor, DT , has to be invented by the analyst. DT
has nothing at all to do with the real world; it is a figment of the calculation.
It is, therefore, afundamental error to use DT on the right-hand side of a
rate or auxiliary equation. Those equations describe information and ac-
tions in the real system; the real system does not contain DT, so the
equations describing the system's workings must not contain DT. There are
two exceptions to this rule, which we shall encounter in Chapter 5.
SUMMARY
This chapter has necessarily been rather long in order to cover the essential
foundations for much further work. We have explained the fundamental
ideas about how system dynamics simulation works and showed the main
equation types, such as delays and smoothing. These concepts were illus-
trated in a simple model. That model was also used to test the effects of
different policies on the behaviour of the system and we emphasized that
policy can have significant effects on system performance; the effects

sometimes being contrary to intuition and common sense. Finally, we
have discussed the selection of a suitable value of DT and considered its
significance.
CHAPTER 5
System dynamics modelling

techniques
INTRODUCTION
In the previous chapter we built a very simple model to demonstrate the

principles of simulation, to show the type of output which a system dynam-
ics model produces and to introduce the idea of counter-intuitive behav-
iour , which is very important in policy design. In Chapter 3, however, we
saw that influence diagrams can be much more complicated than the simple
diagram for the city ruler in Fig. 4.5 so, if system dynamics is to be useful for
quantitative modelling, it must be capable of representing the rather com-
plex problems which can arise in real systems. This chapter therefore covers
some standard techniques and tips for modelling to act as a bridge from
Chapter 3's influence diagrams to the models we shall develop for some of
those problems in the next chapter. For example , we shall find out how to
write the equation which produces the noisy behaviour of FINRA in Fig.
4.11.
At this stage, you should go through this chapter with the aim of finding
out what it contains so that, when in a later chapter you see a CLIP function
for instance, you will know where to find out more about it. In a sense , this
chapter is a very abbreviated user manual for system dynamics software. In
most cases, the forms for functions are stipulated without using system
dynamics conventions, and then illustrated with the conventions. The chap-
ter covers basic features which are widely used; more advanced topics are
treated in Chapter 10.
It was stated earlier, but bears repetition, that the secrets of success in
system dynamics are first to build a good influence diagram and then to
make careful, painstaking and imaginative use of whichever of the several
system dynamics software packages is available. Success does not really lie
in using any particular package though they differ considerably in their
capabilities.
The equation forms used in this chapter are as close as possible to
traditional system dynamics formats , deriving mainly from the DYNAMO
compiler, which was the first to be developed. Some packages differ quite
considerably from that 'standard' and you must consult the manual for the
112 System dynamics modelling techniques
package you are using. This chapter focuses on ideas and good practice, not
on the use of a particular package. As mentioned earlier, this book was
written using the COSMIC and COSMOS packages but it must be empha-
sized that the book is not about COSMIC modelling; it is about system
dynamics modelling.
Throughout this chapter, features of system dynamics modelling will be
illustrated by a few equations. The reader should practise setting up these
examples as complete models and runn ing them to demonstrate the behav-
iour discussed. Bearing in mind the requirements of your own software, use
Fig. 4.6 as a guide to good practice, and follow the advice to print out all
variables and study the results, as well as looking at the graphs of dynamic
behaviour on the screen. A value of DT=0.25 will be suitable for all these
examples. Do not neglect these simple exercises. The practice is vital in
gaining confidence in using your software and even more so in developing
the trick of thinking in terms of the passage of time.
TIME
As was mentioned in Chapter 4, all the system dynamics packages contain

an inbuilt variable called TIME. An extra equation is created automatically
which, if it were written as a level, would be:
L TIME.K =TIME.J + DT
This is, therefore, a perfectly ordinary level equation which acts as a
memory for time in the model. The equation is hidden from the user who
cannot amend it in any way. TIME's meaning is defined as shown in line 46
of Fig. 4.6.
The units in which TIME is measured have to be chosen to suit the
problem. For the predator/prey model, months might be appropriate; for
the decline of the Maya, certainly years, perhaps decades ; for the combat
model , probably hours . DT, the time-step, must have the same units as
TIME. In most of the illustrations in the chapter, TIME will be expressed in
weeks.
Like all levels, TIME requires an initial value which is automatically set
to
NTIME=O
though the user can override that and initialize TIME by including in his
model an equation such as
N TIME = 1990
to simulate time starting on a given date .
The duration for which the model is to run is the maximum value for
TIME and is set by, for example ,
A time-dependent process - the RAMP function 113
C LENGTH = 100
which would simulate 100 weeks starting from 0 if TIME has not been
initialized.
Although the equation for TIME cannot be altered, TIME can be used in
equations just like any other variable. It has a time label of .K unless it is
used on the right-hand side of a level equation, in which case its label is .J.
A TIME-DEPENDENT PROCESS - THE RAMP FUNCTION
In the model for Fig. 4.7 and 4.11 we used the STEP function to generate a
pattern of food inflow which depended on the passage of time. Another
useful time-dependent process is the RAMP function, which takes the
form:
R FLOW .KL=RAMP(SLP,START)
C SLP=2
CSTART=5
D FLOW=(UNIT/WEEK) A FLOW VARIABLE IN A MODEL
D SLP=(UNIT/WEEK) A SLOPE FOR FLOW
D START=(WEEK) THE TIME POINT AT WHICH THE SLOPE
X STARTS
The effect of RAMP is that the rate variable, FLOW , will start at 0 at
TIME=O. When the inbuilt clock of TIME reaches START, FLOW will
increase at a steady rate of SLP per week, so that, by TIME=20, the value
of FLOW will be 30. X is the label for a continuation line.
One could also write
R FLOW.KL=BASE+RAMP(SLP,START)
Run a model to find the effects of
R FLOW .KL=RAMP(SLP1,STARTl)+RAMP(SLP2,STAR2)
C SLP1=2
C STARTl=5
C SLP2=-4
CSTART2=20
Also, find what happens if SLP2=-2.
Predict in your head the effects of
R FLOW.KL=STEP(STEPHT,STPTM)+ RAMP(SLP,START)
C STEPHT=20
C STEPTM=10
C SLP=3
CSTART=30
and test your predictions with a model. What happens if START=5?

The point of these questions is to show that all the standard functions, not
just STEP and RAMP, can be combined in any sequence and to any degree
of complexity. Complexity for its own sake is not a virtue in a model, but
users are restricted only by their own understanding of a problem. There
are no limits in the software on what can be modelled.
In fact, there is no reason why factors such as SLP, or any of the constants
used later in this chapter, have to be constant. They can be variables and
might themselves be calculated from system dynamics functions, as we shall
see later.
In the examples above, FLOW is a rate variable, but we could equally
well have modelled an auxiliary variable, VAR, by the same method. How-
ever, the dimensional arguments and the form of the first fundamental
equation of system dynamics explain why
L LEV.K=RAMP(SLP,START)
would not be legitimate.
RAMP could, however, be used directly for the rates in a level equation,
perhaps to show the effects of rabbits immigrating into an area:
L RABPOP.K=RABPOP.J+DT*(RAMP(IMIG,IMTIME)+ . . .
D RABPOP=(RABBITS) ACTUAL POPULATION OF RABBITS
X IN THE AREA
D IMIG=(RABBITS/MONTH) RATE AT WHICH RABBITS
X IMMIGRATE INTO THE AREA
though it would be better practice to define a separate variable which could
be printed and plotted to make it easier to check that the model is working
correctly:
L RABPOP.K=RABPOP.J+DT*(IMRATE.JK+ . . .
R IMRATE.KL=RAMP(IMIG,IMTIME)
MORE TIME-DEPENDENT PROCESSES-

THE SINE AND COSINE FUNCTIONS
The standard trigonometric functions can be used to represent periodic

time-dependent behaviour, usually in a model's exogenous input. Thus
R NO.KL=BASE+AMP*SIN(6.283*TIME.KlPERD)
C BASE=100
C AMP=30
CPERD=25
D NO=(WMS/WEEK) INFLOW OF NEW ORDERS TO WASHING
X MACHINE PROBLEM OF FIGURE 2.7
Sudden changes - the PULSE function 115
D BASE=(WMSIWEEK) BASE LEVEL FOR NEW ORDER

X INFLOW
D AMP=(WMSIWEEK) AMPLITUDE OF OSCILLATION IN
X NEW ORDER INFLOW
D PERD=(WEEK) PERIOD OF OSCILLATION IN NEW ORDER
X INFLOW
The number 6.283 is the value of 21T so this equation would make NO start
at 100 and oscillate between 130 and 70 with a period of 25 weeks. If some
number other than 6.283 is entered, the sine function will still work, but, in
this case, the period will not be 25 weeks. Note the use ofTIME.K within an
equation.
SUDDEN CHANGES - THE PULSE FUNCTIONl
What would happen to the washing machine firm if its usual pattern of
orders fluctuated as above, but, in addition, it suddenly received a huge
order for 500 washing machines? This requires the use of the PULSE
function, which takes the form
V= PULSE(HEIGHT,FRST,INTVL)
where V is a variable or a term in an equation, HEIGHT is the height of the
PULSE, which lasts for 1 DT, FRST is the value of TIME at which the first
pulse occurs and INTVL is the interval between successive pulses; FRST =
10 and INTVL=20 would make pulses happen at TIME=10, 30, 50 and so
on until LENGTH is reached. If we wrote
R NO.KL= ... + PULSE(OSIZE,FRST,LENGTH)
C OSIZE=500
C FRST=50
D OSIZE=(WMS) SIZE OF LARGE ORDER
where ... represents the previous equation for NO, there would be one
pulse at TIME=50 and then no more, as the second would be called
for at TIME=50 + LENGTH and the model would have stopped at
TIME=LENGTH.
However, this would not produce the required effect, because the
PULSE lasts only for 1 DT and , as DT=0.25, the effect would be a height
of 500 multiplied by 0.25 weeks, or 125 in total. The correct form would be
R NO.KL= . . . + PULSE(OSIZE/DT,FRST,LENGTH)
which makes OSIZE, which is measured in washing machines because it is
a single large order, appear as a rate of inflow of 2000 washing machines per
week for 0.25 weeks, or 500 in all, as required'. This is an excellent example
1 Be very careful when using the PULSE function . Some pack ages tre at it in different ways, so
check the user manual for the package.
2 Bold type cannot, of course. be used when entering this equ ation into a model.
of the use of dimensional analysis to get the correct equation. The first
attempt required washing machines per week, which are the dimensions of
the . . . part of the equation, to be added to washing machines, which are the
units for OSIZE.
This use of DT on the right-hand side of a rate or auxiliary equation
is one of the two exceptions mentioned on p. 109 to the rule prohibiting
that. The other is to handle non-negativity constraints, which are discussed
below.
As well as being used to inject quantities into a model, PULSE can also
be used to make sudden changes in endogenous variables. Suppose that one
wished to run a model with time measured in months, for a LENGTH of 10
years with DT set at 0.25. Some processes in the system, such as financial
budgeting, take place annually, so it is necessary to record the time within
a year and set that record back to zero at the end of each year.
The following equations would be appropriate:
L TYEARK=TYEARJ + DT*(l- PULSE(YLENIDT,YLEN,YLEN))
NTYEAR=O
C YLEN=12
D TYEAR=(MONTH) TIME WITHIN A YEAR
D YLEN=(MONTH) NUMBER OF MONTHS IN A YEAR
TYEAR will grow from 0 to 12 in the first year and then the pulse will
subtract a year from TYEAR during the first DT of year 2. During that DT,
however, the first DT of year 2 will be accumulated so that TYEAR will fall
to the value of DT at the end of the first DT of each successive year. In
short , TYEAR will only be zero when TIME=O . To make TYEAR be zero
at the end of the year the first equation would have to be
L TYEAR.K=TYEARJ + DT*(l- PULSE(YLENIDT,
X YLEN-DT,YLEN))
The reader should experiment!
INTERMITTENT INFORMATION - THE SAMPLE FUNCTION
It is often necessary to represent information being of interest or available

only at intermittent times. For example, whether or not to cull faxes might
be decided at intervals of 6 months, with the decision then being in force for
the next 6 months . This intermittent measurement of the fox population can
be modelled by:
A OBFOX.K=SAMPLE(FOXPOP.K,INTVL,INIT)
where OBFOX means the number of faxes observed when the sample was
taken, INTVL is the interval between samples, which are taken when
TIME=INTVL, TIME=2*INTVL and so on, and INIT is the value of
Random effects - the NOISE and NORMD functions 117
OBFOX until the first sample is taken, in other words the value which
OBFOX will have from TIME=O to TIME=INTVL. In order to get the
initial value right it is usually possible to write:
A OBFOX.K=SAMPLE(FOXPOP.K,INTVL,FOXPOP.K)
in which case the initial value of OBFOX will be set to the initial value of
FOXPOP. Note the use of a variable as one of the inputs to this function .
Obviously, OBFOX, FOXPOP and INIT must all have dimensions of
[FOXES] and INTVL must have the same dimensions as TIME.
RANDOM EFFECTS - THE NOISE AND NORMD FUNCTIONS
Random effects are often included in system dynamics models . There are
two functions: NOISE, which produces values uniformly distributed be-
tween -0.5 and 0.5, and NORMD, which samples from a Normal
(Gaussian) distribution of given mean and standard deviation (some pack-
ages use different names for this function). The principles for using both are
the same , so we shall discuss only the NOISE function.
It is tempting to write and, indeed, one frequently sees in system dynam-
ics models, an equation such as
A VAR.K=NOISE(SEED)
C SEED=99 (for instance)
A VAR=(UNIT) A VARIABLE WHICH VARIES UNIFORMLY
X BETWEEN 0.5 AND -0.5
D SEED=(l) A NUMBER TO SEED THE RANDOM SAMPLING
in which VAR is a random variable and the significance of SEED is that,
every time it is set to a given value, the same sequence of random numbers
will be generated. SEED is a simple number which is regarded as
dimensionless, so its dimensions, or lack of them , are denoted by (1).
The equation above is, however , fundamentally wrong. The reason is
that, in system dynamics models, every variable is calculated every DT, so
the equation says that a new random sample is to be taken every DT. If DT
was halved , random samples would be taken twice as frequently. However,
as we discussed in the last chapter, DT is an invention of the modeller in
order to get the model to run on the computer, and it has nothing whatever
to do with the nature of the real system. In this case, we must take samples
at the same frequency that nature in the real world generates randomness,
and we must ensure that the randomness is of the same magnitude as
nature's. This requires us to combine two of the system dynamics functions
and put
A VAR.K=SAMPLE(SCALE*NOISE(SEED),NFREQ,
X SCALE*NOISE(SEED))
C SCALE=O.l (for example)
D SCALE=(l) A MULTIPLIER TO MAKE THE SAMPLES HAVE

X THE SAME RANGE AS NATURAL VARIATION
D NFREQ=(WEEK) THE FREQUENCY WITH WHICH
X RANDOMNESS OCCURS IN THE REAL
X WORLD
This will produce a fresh sample every FREQ weeks and SCALE will force
the randomness to lie in nature's range, whatever it is, not in NOISE's of 0.5
to -0.5. In this example, the random variable, VAR, will be between 0.1 x
0.5 and 0.1 x -0.5, or 0.05 to -0.05. The use of SCALE and NOISE in the
third argument of SAMPLE ensures that a suitable random value is used
between 0 and TIME=FREQ.
To illustrate NOISE more specifically,suppose one wished to model the
random effects of minor breakdowns on production capacity. Clear think-
ing is required because, when the sampling of NOISE produced positive
values, one would be close to saying that production capacity could be
increased by random effects. A possible approach would write
A PCAP.K=NCAP*RANEFF.K
A RANEFF.K=(1-SCALE*0.5+SAMPLE(SCALE*NOISE(SEED),
X FREQ,SCALE*NOISE(SEED))
D PCAP=(UNIT/WEEK) CURRENT PRODUCTION CAPACITY
D NCAP=(UNIT/WEEK) NOMINAL OR MAXIMUM
X PRODUCTION CAPACITY
D RANEFF=(l) A MULTIPLIER TO REDUCE CAPACITY TO
X REFLECT THE RANDOM EFFECTS OF MINOR
X BREAKDOWNS
The effect of these equations is that a sampling which produces the result
0.5 will be exactly cancelled by the SCALE*0.5 from which the random
values are subtracted, and there will be no increase in production above the
maximum capacity. If the sampling produces any other value, production
capacity will be correspondingly reduced.
Note that SCALE is used both within the SAMPLE and outside it to
ensure that, even if this parameter is subsequently changed to test
the robustness of the system policies against an increased magnitude
of randomness, the equation will still be correct. Note also the use
of RANEFF as a separate variable which can be printed or plotted to
ensure that it never exceeds 1, a useful check that the model is working
correctly.
It would be bad practice to write
A PCAP.K=NCAP*(l-SCALE*O.5+SAMPLE(SCALE*NOISE(SEED),
X FREQ,SCALE*NOISE(SEED))
precisely because it would require more effort to see that what is happening
Logical operations - the CLIP function 119
in the model is correct. The greater the visibility of a model 's workings the
more one can be confident in it.
LOGICAL OPERATIONS - THE CLIP FUNCTION
The example for the SAMPLE function implies that culling of foxes might
take place if OBFOX was greater than some target value but not otherwise .
This is an example of a logical choice which, in the original system dynamics
packages , requires the use of a function with the strange name of CLIP.
This takes the form:
V=CLIP(A,B,C,D)
in which V=A if C~D and V=B if C<D. In dimensional terms , [V]=[A]

=[B] and [C]=[D], in other words, C and D must be alike in order to be
validly compared and A and B must be consistent with the result, V. Notice
carefully that the answer A is produced when C is greater than or equal to
D.
°
To apply this to the fox problem we could define a 'culling switch' which
is 1 when culling can take place and when it is forbidden . That could be
done by:
A CSWTCH .K=CLIP(l ,O,OBFOX.K,CRITF.K)

D CSWTCH=(l) A SWITCH TO INDICATE WHETHER OR NOT
X CULLING OF FOXES CAN TAKE PLACE
D OBFOX=(FOXES) OBSERVED NUMBER OF FOXES
D CRITF=(FOXES) CRITICAL NUMBER OF FOXES AT OR
X ABOVE WHICH CULLING IS REQUIRED
The foregoing equation could be linked with another:
R CULLR.KL= DCULLR.K*CSWTCH .K
D CULLR=(FOXES/MONTH) ACTUAL RATE AT WHICH
X FOXES ARE CULLED
D DCULLR=(FOXES/MONTH) DESIRED RATE OF CULLING
X FOXES
The CLIP function produces the result A if C~ D so how does one handle
the case that V=A if C>D, a so-called strict inequality? One approach
would be to write:
A CSWTCH .K= CLIP(O,l ,CRITF.K,OBFOX.K)
which will give the answer °
if CRITF.K ~OBFOX.K and, therefore,
the answer 1 if CRITF.K< OBFOX.K, which is the same as
OBFOX.K ~CRITF.K. Anoth er approach would use
A CSWTCH.K= CLIP(l ,O,OBFOX.K,CRITF.K+ EPS)

C EPS=O.01
D EPS=(FOXES) A SMALL NUMBER OF FOXES
which makes the CLIP produce 1 as soon as OBFOX.K equals or exceeds
CRITF.K plus a small amount. The value of EPS need not be 0.01 but can
be chosen to suit the problem .
It may seem confusing to have two ways of tackling a problem, but
the point is to show that the system dynamics packages are almost in-
finitely flexible and there can be many ways of achieving the same
result. Which one is used is partly a matter of personal style. In such
instances, the author prefers the EPS method, as it is consistent with the
standard use of CLIP and consistent practice is usually a good way of
avoiding errors.
Logic requiring two conditions to be met can be modelled using two
CLIPs3. If, for example, V=A if C~D AND E~F, to represent the foxes
being culled if there are too many of them and the fox breeding season is
over, then one could put:
A CSWTCH.K= CLIP(l,O,OBFOX.K,CRITF.K) *
X CLIP(l,O,TYEAR.K,BRDTIM)
D TYEAR=(MONTH) A VARIABLE RECORDING THE TIME
X WITHIN A YEAR, NOT THE SAME AS
X TIME WITHIN THE MODEL
D BRDTIM=(MONTH) THE TIME WITHIN THE YEAR AFTER
X WHICH FOXES STOP BREEDING
If both conditions are satisfied, then both CLIPs will produce 1, and the
product of the two values will set CSWTCH to 1. If however, one condition
or other is not met, one of the CLIPs will produce the value 0, which will
negate the value being produced by the other CLIP. It would, however, be
better practice to write
A CSWTCH.K=FOXCON.K*YEARCON.K
A FOXCON.K=CLIP(l,O,OBFOX.K,CRITF.K)
A YEARCON.K=CLIP(l,O,TYEAR.K,BRDTIM)
D FOXCON=(l) SWITCH TO INDICATE THAT FOX NUMBERS
X REQUIRE CULLING
D YEARCON=(l) SWITCH TO SHOW THAT CULLING IS
X PERMISSIBLE
This would allow FOXCON and YEARCON to be printed out so that one
could check that the model was working correctly. Obviously, one could
3 COSMIC is one of the system dynamics packages which allow for more complex logic.
In this case COSMIC has single functions which are the equivalent of two CLIPs com-
bined into one. Other packages allow logic to be entered in the form 'IF x THEN YELSE z'
etc.
Limiting functions - MAX and MIN 121
also check that by looking at the values of the arguments to the CLIPs and
making mental calculations, but that is more prone to error, and the less
chance of error in a model the more one can be confident that it is doing the
same things as the real system and for the same reasons.
Logic involving V=A if C~D OR E~F can also be handled using two
CLIPs:
A VARK=CLIP(l ,O,C.K,D.K)+CLIP(l,O,E.K,F.K)
or, preferably, for the reasons given above ,
A VARK=VAR1.K+VAR2.K
A VAR1.K=CLIP(l,O,C.K,D.K)
A VAR2.K=CLIP(1,O,E.K,F.K)
This is correct unless both C~ D and E~ F in which case both CLIPs

produce the value of 1 and VAR will be equal to 2. One way round that
would be to write:
A VARK=VAR1.K+VAR2.K*(1-VAR1.K)
with VARI and VAR2 as before. In this case, if VARI = 1 and VAR2=O,
then VAR=l .IfVARl =O and VAR2=1, then VAR=l, but if both VARI
and VAR2 are 1 then VAR is still 1, as required. As we shall see later, there
is another way of achieving this result.
This example also illustrates an aspect of writing system dynamics models
that one needs to understand at an early stage. In a normal programming
language, such as BASIC or Pascal, the programmer will arrange for logical
paths that are followed in some cases, with different paths being taken in
other cases; parts of the program are simply not used in some circum-
stances. System dynamics modelling is different from computer program-
ming in that all the system dynamics languages work by calculating all the
variables every DT. One cannot bypass VARI and use VAR2 , or vice
versa; both are always 'alive ' and the modeller must arrange to switch off
the one not required. This is a habit which quickly becomes second nature,
but it requires a little practice and careful debugging of the model in one's
early stages of learning. There is no substitute for printing out all the
variables in a model, as was done in Fig. 4.6, and carefully studying the
printout to check what is happening to the variables as time passes in the
model.
LIMITING FUNCTIONS - MAX AND MIN
The maximum and minimum functions have their usual form of

V=MAX(A,B) and V=MIN(A,B)
where at least one of A and B must be a variable if the use of MAX or MIN
is to have any point. A triple maximum cannot be written as
V=MAX(A,B,C)
Instead one must write
v = MAX(A,MAX(B,C))
MAX and MIN can be combined, if desired, for example
V=MAX(A,MIN(B ,C))
to any desired degree of complexity .
The previous example of VAR1 and VAR2 can be handled easily by
writing
A VARK = MIN (VAR1.K + VAR2.K,1)
The purpose of showing the more complicated method was to emphasize
that, in system dynamics languages, there are always several ways of solving
any given problem.
MAX and MIN, like all the system dynamics functions , can be 'nested' in
this way, sometimes up to 9 deep, depending on the package . Nesting is,
generally, very inadvisable as it makes testing the model's operations for
correctness more difficult. In proper system dynamics notation, and assum-
ing that all the variables are auxiliaries, one should write
A VARK=MAX(A.K,DUMMY.K)
A DUMMY.K=MAX(B.K,C.K)
so that DUMMY can be printed or plotted to make it easier to see that what
is happening in the model is happening for the right reasons.
MAX and MIN should be used only when the logic of the model requires
them, and never to cover up strange behaviour. For example, one might
wish to model Market Share, MS, as depending on a rate of growth, MSGR
To do so requires a level equation
L MS.K= MS.J+ DT *MSGRJK
It is evident that MS can never exceed 1, and if the model produces a value
of more than 1 there is a temptation to write
L MS.K=MIN(l,MS.J+DT*MSGRJK)
Most system dynamics packages would accept this, though they might give
a message that this is an unusual format for a level equation, but the
equation is actually wrong. The fault lies in the equation for MSGR, which
is allowing MS to grow past its limit, and using a MIN in this way is simply
covering up a fault somewhere else in the model, which is hardly good
practice.
Non-negativity constraints 123
NON-NEGATIVITY CONSTRAINTS
In the management consulting firm's problem in Fig. 3.1 and in the combat
model of Fig. 3.7 we encountered non-negativity constraints , though on p.
68 we make it clear that these are two rather different cases.
In the first case, the consulting firm determines the Desired Trainee
Recruitment Rate, DTCRR If this is positive, it indicates that trainees
should be recruited, while a negative value would indicate that trainees
should be dismissed. However, the firm has a policy of never dismissing
people , so we wish to suppress the negative option. That requires
R TCRRKL= MAX(O,DTCRRK)
D TCRR=(PEOPLE/MONTH) ACTUAL TRAINEE
X RECRUITMENT RATE
D DTCRR=(PEOPLE/MONTH) DESIRED TRAINEE
X RECRUITMENT RATE
a perfect instance of a MAX being used where the logic of the problem
requires it.
In the second case, once combat starts, the Red commander commits
reserves as fast as he can, as long as he has reserves left to commit. How
fast he can commit them depends solely on the transport capacity available
to him. That can be modelled by using the following three groups of
equations:
L RRES.K = RRESJ + DT*( - RRCRJK)
N RRES = IRRES
C IRRES= 50000
D RRES=(RMEN) RED RESERVES REMAINING
D RRCR=(RMEN/HOUR) RATE OF COMMITTING RED
X RESERVES
D IRRES=(RMEN) INITIAL RED RESERVES
This level simply accounts for the remaining reserves as they are used up.
Note the use of (- RRCRJK) to represent the drain; most packages will not
allow - DT*RRCRJK. It is good practice to set up initial values by using a
constant as shown. The value of IRRES can be changed in a later run , which
would be impossible if one had used N RRES=50000. Note also the use of
RMEN, for Red men, rather than just MEN. It is always good practice to
differentiate one type of unit very clearly from another, such as Blue men.
A WSWITCH.K=STEP(l ,WTIME)
C WTIME=5
D WSWITCH=(l) A SWITCH TO DENOTE THE START OF
COMBAT
D WTIME=(HOUR) TIME AT WHICH COMBAT STARTS
124 System dynam ics modelling techniques
This is a simple use of STEP to ensure that nothing happens in the first 5
hours of the model run; a useful trick to make sure that the model is set up
correctly.
R RRCR.KL= MIN(RRES.K/DT,RTCAP)*WSWITCH.K
C RTCAP=6000
D RTCAP=(RMEN/HOUR) RED TRANSPORT CAPACITY
The MIN term in the equation for RRCR is prevented by WSWITCH from
having any effect until TIME=5; another instance of switching something
off when it is not required. Within the MIN, the second factor, RTCAP,
is obvious; it is the first factor, RRES.K/DT which represents the non-
negativity constraint. To see how it works, let DT=l for simplicity.
At TIME=5 the combat starts and, since RRES is initially 50000, RTCAP
will ensure that Red can only commit reserves at the rate of 6000 men
per hour. This can go on for 8 hours, by which time only 2000 men will
remain. If there was no non-negativity constraint, a commitment rate of
6000 per hour for another DT would leave Red with -4000 men remaining ;
a clear absurdity . However , at that point , RRES.K/DT will be 2000, so
the last 2000 men will be committed evenly during one more DT, at which
point there will be exactly 0 men left and reserve commitment will have to
stop.
In the above equations, IRRES, RTCAP and WSWITCH are all num-
bers, but note that IRRES and RTCAP have dimensions associated with
real quantities, whereas WSWITCH is dimensionless because it represents
an event occurring rather than something tangible.
Using DT on the right-hand side of a rate equation to model a non-
negativity constraint is the second of the two exceptions to the prohibition
against doing so. Whenever the modeller writes DT on the right-hand side
of a rate or auxiliary equation it is essential to stop and work out whether
one is using one of the two valid exceptions to the rule, the reasons for
which are given on p. 109, or whether one is making a fundamental and
catastrophic error of system dynamics logic. Notice, for instance, that DT is
not used in the consulting company's non-negativity constraint.
CHOOSING A SUITABLE DELAY FUNCTION
In Figs. 4.8 and 4.9 we discussed delays and showed that they produced
different dynamic responses to a STEP input. We also stated that
there were other types of delay and, in particular, tested the perfect or
pipeline delay which exactly mirrors the input after a time lag. How,
though, does one know which delay to use in a given case and, indeed ,
how much difference it will make if one uses one delay type rather than
another?
The best way to understand any of the system dynamics modelling tech-
niques is to build a model and experiment, and this is no exception . Figure
5.1 is a model to show four different delay types against two different input
patterns. This model is on the disk as FIG5-2.COS, in COSMIC format,
though the reader should have little difficulty in converting it and, as ever,
the emphasis is on the principles, not on the software.
In Fig. 5.1 the first two lines activate the model mapping which is de-
scribed below and initiate automatic dimensional analysis. The driving
force is in lines 8 to 18 and the reader will see a switch, SW1, to choose
between a PULSE input and a noisy pattern which is very close to that used
in Fig. 4.11, the equation being explained in detail towards the end of this
chapter. The switch is changed in lines 84 and 85 to run the model twice.
This is a traditional system dynamics technique which is useful when one
wants a hard copy of the output. All the system dynamics packages also
allow one to make changes like this interactively.
The output for four different cases is modelled in lines 31, 39, 47 and 56-
57. The four cases are first-order, third-order (recall Fig. 4.8), pipeline and
'sixth-order' delays. The last of these is built up using two cascaded third-
order delays, each of which has half the total delay. This requires the use of
a dummy rate, DOUT4 in line 56 as an intermediary between the two
DELAY3s. Cascading like this can produce any delay order. The reader
should amend FIG5-2.COS to include a ninth-order delay.
For each delay we also calculate the contents and the cumulative output,
as in lines 32 to 35 for the DELAYl. Line 48 is a special COSMIC require-
ment to load the pipeline initially as one wishes to have it; in this case we
want it to be empty.
To check that the delays are functioning correctly, mass balances are
computed in lines 65 to 68. Such balance checks are a most useful method
of debugging models involving complex flows, though they are not strictly
necessary here. Note that all the variables are printed and a selection are
plotted.
Although this is a simple problem to help us in our study of delays,
you may find it helpful to adopt the conventions of Fig. 5.1 when using your
own package, as they represent much of what the author regards as good
system dynamics modelling practice, in particular the heavy emphasis on
generating output to assist in verifying that the model is working correctly
and the full documentation of every variable and parameter.
The output for the two cases is shown in Fig. 5.2.
The PULSE happens at TIME =10 and the graphs for INRAT and 0 UT3
go off the page, as both reach a value of 200/DT. However, the exact match
between the two, 10 days apart, is easily seen. One can also see the imme-
diate response and long tail of DELAYl and the progressively sharper
humps of the third- and sixth-order reactions . The contents of the respec-
tive delays in the lower half of the page show broadly similar patterns,
noting especially the square form of CONT3 for the pipeline, the amount
being 200 units, exactly as one would expect.
126 System dynamics m odelling techniques
°1
map
dim
2 * Figure 5.2 Delay Behaviour
3 note
4 note file named fig5-5.cos
5 note
6 note input patterns
7 note
8 r inrat .kl= sw 1* pulse(heightJdt,f rst,i ntvll+
9 x (1-sw11 * step(base+ am p* sin(6.283* (t im e.k-151/perdl
10 x * clip(1,0,ti m e.k,151 *sample(noise(7l.nperd ,01.1 01
11 csw1=1
12 c height =200
13 c frst= 10
14 n intvl=length
15 c nperd =2
16 c base =40
17 c amp =20
18 c perd =20
19 note
20 note cumu lative input for balance check
21 note
22 I cumin .k=cumin .j + dt* inr at.j k
23 n cumin =O
24 note
25 note the delay magnitude
26 note
27 c del=10
28 note
29 note output rate for first order delay
30 note
31 r out1.kl =delay1( inrat.jk,dell
32 I cont1.k =cont1.j + dt * (inrat.jk-out 1.jkl
33 n cont1 =0
34 I cumout1.k =cumout1.j + dt*o ut 1.jk
35 n cumout1 =0
36 note
37 note output rate for second order delay
38 note
39 r out2 .kl = delay3(inr at .jk,dell
40 I cont2.k =cont2.j +dt*(inrat.jk-out2.jkl
41 n cont2 =0
42 I cumout2 .k=cumout2 .j+dt *out2.jk
Fig.5.1 Model for delay demonstration. 'Figure 5.2' appea rs in line 2 because this
line is used to generate the caption for Fig. 5.2.
43 n cumout2=0
44 note
45 note output rate for pipeline delay
46 note
47 r out3.kl=dlpipe(tvals,inrat.jk,del)
48 t tvals=40*0
49 I cont3.k=cont3.j+dt*(inrat.jk-out3.jk)
50 n cont3=0
51 I cumout3.k=cumout3.j+dt*out3.jk
52 n cumout3=0
53 note
54 note output rate for sixth order delay
55 note
56 r dout4.kl = delay30nrat.jk,del/2)
57 r out4.kl=delay3(dout4.jk,deI/2)
58 I cont4.k=cont4.j+dt*Onrat.jk-out4.jk)
59 n cont4=0
60 I cumout4.k=cumout4.j+dt*out4.jk
61 n cumout4=0
62 note
63 note mass balance checks
64 note
65 a mchck1.k=cumin.k-cumout1 .k-cont1.k
66 a mchck2.k=cumin.k-cumout2.k-cont2.k
69 note
70 note model control and output
71 note
72 c dt=0.25
73 c length=50
74 c prtper=10
75 c pltper=0.25
76 print 1linrat, cumin
77 print 2)out1 ,cumout1 ,mchck1
78 print 3)out2,cumout2,mchck2
81 plot inrat=a,out1 =b,out2=c,out3=d,out4=e(0,60)
82 plot cont1 =a,cont2=b,cont3=c,cont4=d(0,600)
83 run Delayed Response to a Pulse Input
84 c sw1=0
85 run Delayed Response to a Noisy Input
Fig. 5.1 Continued

86 note
87 note definitions of variables
88 note
89 d amp={units/day) amplitude of sine wave in input
90 x rate
91 d base={units/day) base value for input rate
92 d cont1 =(units) contents within the first order
93 x delay
94 d cont2={units) contents within the second order
95 x delay
96 d cont3={units) contents within the pipeline delay
97 d cont4={units) contents within the sixth order
98 x delay
99 d cumin={units) cumulative input
100 d cumout1 =(units) cumulative output for first order
101 x delay
102 d cumout2={un its) cumulative output for third order
103 x delay
104 d cumout3={units) cumulative output for pipeline
105 x delay
106 d cumout4={units) cumulative output for sixth
107 x order delay
108 d del={day) magnitude of delay
109 d dout4={units/day) dummy output for first delay3
110 x in sixth order delay
111 d dt={day) step size in simulation
112 d frst={day) time of first pulse in input rate
113 d height={units) size of pulse in input rate
114 d inrat={units/day) input rate to various delays
115 d intvl={day) interval between pulses in input
116 d length={day) duration of simulation
117 d mchck1 =(units) mass balance check for first
118 x order delay
119 d mchck2={units) mass balance check for second
120 x order delay
121 d mchck3={units) mass balance check for pipeline
122 x delay
123 d mchck4={units) mass balance check for sixth
124 x order delay
125 d nperd={day) period of random noise in input
126 x rate
127 d out1 =(units/day) output rate from first order
128 x delay
Fig. 5.1 Continued

129 d out2=(units/day) output rate from second order

130 x delay
131 d out3=(units/day) output rate from pipeline
132 x delay
133 d out4=(un its/day) output rate from sixth order
134 x delay
135 d perd=(day) period of underlying sine wave in
136 x input rate
137 d pltper=(day) plotting interval
138 d prtper=(day) printing interval
139 d sw1=(1) switch to change input rate patterns
140 d time=(day) simulated time
141 d tvals=(units/day) table to load pipeline initial
142 x values
143 +
No syntax errors detected
Fig. 5.1 Continued
The conclusion from Fig. S.2(a) and (b) is that the choice of delay order
matters quite strongly when one is dealing with very sharp inputs , such as
PULSEs or a STEP. The difference between DELAY3 and the cascaded
sixth-order delay is, perhaps, not all that large; probably well within the
uncertainties of the data in most managed systems. One can, perhaps,
conclude that for practical purposes one can distinguish between three main
possibilities: the output being an exact copy of the input, in which case a
pipeline delay is appropriate; the case where the input has an immediate
effect and then a long tail, DELAYl ; and an intermediate case where there
is a delayed response and then a spread of inflow, in which case DELAY3
will usually be appropriate. It is very tempting to use the pipeline under
some illusion of accuracy, but such cases are rare in managed systems,
dispersed delays being far more usual.
Fig. S.2(c) and (d) shows a more usual case of a continual, though very
unsteady, inflow. Again the pipeline matches it exactly, whereas the three
distributed delays smooth out the input considerably and, indeed, have little
difference between them once the effect of the STEP has been absorbed.
Continuous flows are far more typical than sharp shocks in the modelling of
managed systems. In such cases, the 'accuracy' of a pipeline is largely
illusory, and DELAY3 is a good compromise well within the accuracy of
the data usually available , which is why it is so commonly used in system
dynamics modelling.
It is worth reiterating that all the system dynamics packages have a
BO .
TIME (DAY) snU.ATED TIME
!I'.,'- - LNRAT (UNITS/DAY) INPUT RATE TO YARIOUS DELAYS
I -OUT1
: 1.-._ UT3
ogUT2
(UNITS/DAY)
(UNITS/DAY)
(UNITS/DAY)
OUTPUT RATE FROM FIRST ORDEA
OUTPUT RATE FROM SEClNI ORDER
OUTPUT RATE FROM PIPELINE
(a) L..-··-OUT4 (UNITS/DAY) OUTPUT RATE FROM SIXTH llRIlER
II L· TIME (DAY) SIIOJLATED TIlE

· 1 CgNT1 (UNlTSJ CONTENTS wITHIN TtE FIRST a:IOER
1 -- C NT2 (UNITSI CONTENTS wITHIN TIE SECOND ORDER
c:::
E8~H :~ml gg::~~~~ :n~~= ~ ~~~~I::"'~LAY
(b) DELAYED RESPONSE TO A PULSE INPUT
Fig. 5.2 Delay behaviour.
number of delay functions and it is useful to take the time to experiment

with them . In some cases the syntax differs considerably for the forms used
here , though the principles discussed here are applicable.
NON-LINEARITIES - THE TABLE FUNCfION
Much of the interesting dynamic behaviour of a managed system stems

from the presence of non-linear effects within it. The influence of the
Non-linearities - the TABLE function 131
80 ,
lilt
,' ',II INAAT
TIME (CAVI SIMULATEC rIME
(UNITS/CAV) INPUT RATE TC VARlDUS DELAVS
-- OUT1 (UNITS/CAV) OUTPUT RATE FROM FIRST ORDER
. ~'OUT2 (UNITS/CAV) OUTPUT RATE FROM SECOND ORDER
, L-'-OUT3 (UNITS/OAY) OUTPUT RATE FROM PIPELINE
l..-'·-OUT4 (UNITS/CAV) OUTPUT RATE FROM SIXTH OAOEA (c)
800
700
600
400
o 50,
II L
, ~5 , ~o , ~5, ~o . ~5,
=
TIME (CAVI SIIUJoTED rIME
'II CONn (UNITSI CONTENTS OITHIN THE FIRST ORDER
,L-._
..=.~ g8~H
CONT 4
l:ml g:~~~ :g~~= ~~NE~AY
(UNITS) CONTENTS OITHIN THE SIXTH ORDER
DELAYED RESPONSE TO A NOISY INPUT (d)
Fig. 5.2 Continued
Average Food Intake per Fox (AVIPF) in Fig. 5.3 on the lifetime of foxes
is a good example of the difference between linear and non-linear behav-
iour. Suppose that AVIPF was 10% below the level required for a normal
lifetime (NFRPF in Fig. 5.3), and suppose that would reduce fox lifetime to
5% below normal. If AVIPF fell to 20% below NFRPF then the influence
would be linear if lifetime fell to 10% less than normal, because doubling
the food shortfall had doubled the effect. If the relationship was truly linear ,
then a 100% fall in food supply, complete starvation, would lead to a 50%
reduction in lifetime; clearly absurd, so the relationship must be non-linear.
Setting aside for the moment how one collects the numerical data for the
relationship, we can approach the handling of non-linearities in system
dynamics models on the following lines.
In the first place, the model for Fox Death Rate and D U FE would usually
be written in system dynamics notation as
R FDR.KL=DELAY3(FBR.KL,DLIFE.K+MINLT)
A DLIFE.K=NFLTF*FSRF.K
C NFLTF=some suitable number
C MINLT=some suitable number
D FBR=(FOXES/MONTH) FOX BIRTH RATE. AN EQUATION
X FOR THIS VARIABLE EXISTS
X SOMEWHERE ELSE IN THE MODEL
D FDR=(FOXES/MONTH) FOX DEATH RATE
D DLIFE=(MONTH) CURRENT FOX LIFETIME
D NFLTF=(MONTH) NORMAL FOX LIFETIME IF FOOD
X INTAKE IS AT THE LEVEL REQUIRED
X FOR A NORMAL LIFESPAN
D FSRF=(l) A MULTIPLIER TO REPRESENT THE EFFECTS
X OF AN INADEQUATE FOOD SUPPLY ON FOX
X LIFESPAN.
~ _ --.._-
o
00
- ",• " :, •
' . . - .
"
.5
~ . . .. . . . .. . . . . . . . . . . -. .-
"
~ ,"
.a
", " '\ '
ec "-
c c:::
'n til
"-
.g
...e
.~
;; . . . • • . ' . . . . • • '. . • . . - ! • . • . . • '• • . . • . •
:e.
:g
s
"
.c
I-
N
o
•
' ," . . - . :
• •
",
•
" .
~L...--=== _
0.0 0.2 0.4 0.6 0.8 1.0
FRATIO
The ratio of actual and requiredfood intake rates
Fig. 5.3 A non-linear relationship between food intake and the effect on fox
lifetime.
Non-linearities - the TABLE function 133
D MINLT=(MONTH) THE MINIMUM LIFESPAN OF A NEW-
X BORN FOX CUB IF THERE IS NO FOOD
MINLT is needed because, if DLIFE.K was 0, the DELAY3 function would

attempt to divide by zero, as shown in the equations in Fig. 4.8. MINLT
obviously act as a delay and, for the reasons given on p. 109, DT will have
to be no larger than MINLT/12.
FSRF is a multiplier to scale NFLTF down according to the availability of
food. Such multipliers are very common in system dynamics modelling, as
they give a convenient way of modelling the effects of one variable on
another, especially when the effect is non-linear. How, though , are we to
model FSRF?
On the principle of introducing new variables to break what would
otherwise have been complicated equations into simpler components to
make it easier to see that the model is working correctly, let us define a new
variable, FRATIO:
A FRATIO.K=AVIPF.K/NFRPF
D FRATIO=(l) RATIO OF ACTUAL AND IDEAL FOOD
X SUPPLIES
D AVIPF=(RABBITS/MONTHIFOX) AVERAGE ACTUAL
X INTAKE OF FOOD PER
X FOX
D NFRPF=(RABBITS/MONTHIFOX) AVERAGE FOOD INTAKE
X PER REQUIRED FOR
X NORMAL LIFESPAN
A hypothetical relationship between FRATIO and FSRF is shown in Fig.

5.3. If FRATIO.K is 1 the foxes are fully fed and their lifespan is not
affected , represented by FSRF.K=l on the vertical scale. If FRATIO.K is
zero, however, foxes will die rather quickly, corresponding to FSRF.K
being zero. Similarly, the value of FSRF can be read from the graph for any
value of FRATIO. Notice that the graph is highly non-linear. Small reduc-
tions below 1 in FRATIO have little effect, but the curve drops steeply
until, at low values of FRATIO, fox lifespan would not be a good insurance
risk. System dynamics notation converts such a non-linear curve, which we
assume to be based on data about fox biology, into a set of linear segments
which fit reasonably closely to the actual curve.
Such a piece-wise linear approximation is shown in Fig. 5.4. The dotted
lines are the curve which the software will use to approximate to the true
curve. It is expressed in the notation:
A FSRF.K=TABLE(TFSRF,FRATIO.K,0,1,0.2)
T TFSRF=0/0.024/0.12/0.8/0.98/1.0
D TFSRF=(l) TABLE FOR THE RELATIONSHIP BETWEEN
X FOOD RATIO AND FOOD SUPPLY REDUCTION
~ -.- - . _ - .. "- - -
o00 ,
.
,
"
.5 I , ':
~ I '
I '
, , ',' l .,
.s"
t>ll
c ""0:: . ' I ' ,
'u
:::l
."
'"
c, I
!!
~
....o ' -, - ' ' .
.2 , : I '
:g. , I
:; . I
E , I
"
~ ;;j : :.". ', " : .
~ 0.0
----====--------------
0.2 0.4 0.6 0.8 1.0
FRATlO
The ratio of actual and required food intake rates
Fig. 5.4 A piecewise linea r approximatio n to the non-linear relationship.
x FACTOR WHICH MODELS THE EFFECTS OF

X TH E AVAILABILITY OF FOOD ON THE
X LIF ESPAN OF FOXES
The equation for FSRF means that the value of FSRF at any point in time
will be calculat ed for a given value of FRATlO from a look-up table called
TFSRF. The three numb ers at the end of the equation provide that
FRATlO will range from a minimum of 0 to a maximum of 1.0 in steps of
0.2. Th is requ ires a total of 6 values to be specified, and they are given in the
second line, which uses the type label T for a table, the individual values
being separated by /. Thu s, the first value, 0.0, mean s that when
FRATIO.K =O, FRSRF.K will also be zero. The third value provides that
FRATlO.K=OA will generate a value of 0.12 for FRSRF. If FRATIO
happ ened to be 0.65, say, then FRSRF would be calculated by linear inter-
polati on between the table values of 0.8 and 0.98 which are the TFSRF
values for FRATlO = 0.6 and FRATlO =0. 8, respectively.
Th e 6 numb ers for TFSRF are found by measuring the points at which
the linear segments in Fig. 504 cross the dott ed vertical lines. The numb er
of linear segments is chose n by eye to get a reasonable match between
the linear segments and the real curve. Figure 504 has been deliberately
drawn to be a mediocre match , especially between 004 and 0.6. One might,
indeed, decide to have twice as many segment s to get a bett er fit and the
reader should practi ce setting up such a table, which will require 11 values
in total.
Non -linearities - the TABLE function 135
In the equations above, note that the definition for TFSRF extends to five
lines of text. Writing very complicated equations is bad practice; it is far
better to break up a complex equation into several simpler auxiliaries. Long
definitions, when called for, are very good practice as they make clear to
client and modeller alike exactly what the model 's variables and parameters
mean. This is true whether the client is a paying a fee for a service or is the
academic audience for a book or paper.
The TABLE function has a strict limitation in that, in most system
dynamics software packages , it will produce a fatal error if FRATIO, in this
example, goes below 0.0 or above 1.0, its declared range. A variant on this
is the TABHL function which has exactly the same format but which allows
the argument, FRATIO, to go outside its declared range , using the first or
last of the table values as a horizontal 'tail' for the output, FSRF, if that
happens. Using TABHL is a temptation but is to be avoided unless the logic
of the problem really requires it. In this example, FRATIO cannot logically
be less than 0 or greater than 1, so it would be helpful for thorough model
debugging if the model crashes when that happens.
If a model does crash because a table has gone outside its range , careful
study of the printout of the model 's variables will usually reveal the cause.
There is no substitute for using printed values of model variables in debug-
ging a model. The graphs of a few variables displayed on the screen will
sometimes look right when they are not, because something is going wrong
which is not being seen .
The examples used here to explain the principles of non -linearities have
been expressed in the DYNAMO-type format which is used by at least two
of the other packages. It is essential, however , to check the manual for the
package you are using as the details of setting up TABLE and TABHL
functions, and even their names, differ considerably in some of the pack-
ages. Most packages will offer additional forms for modelling non-
linearities. For example, COSMIC has table types in which the output does
not vary smoothly but steps up or down as the argument changes .
Tables are a most useful feature in system dynamics and can be used very
flexibly. There is, of course , no reason why the output and the argument
both have to lie between 0 and 1; the range of either has to be whatever is
appropriate to the problem. One useful role for a table is to inject into a
model an historical pattern of, say, washing machine orders. In such a case,
the argument would be TIME.K and the values in the table would be the
sales rates in each of the past 48 months, or whatever.
The dimensions for a non-linearity can be understood by writing the
equation in abstract form:
V=TABLE(TABNAM,ARG,LO,HI,INCR)
Recalling that [] means 'the dimensions of', this requires that
[VAR]= [TABNAM] and that [ARG]=[LO]=[HI]=[INCR]. In words, the
table values have to be in the same units as the result, VAR , and the
argument has to have the same dimensions as the three points which specify
its range . However, there is no requirement at all for [VAR] to be the same
as [ARG] . In short, the TABLE function and its relatives can validly act as
a dimensional transformation between one variable and another. The curve
of the relationship stands for the unknown causal mechanism which relates
ARG to VAR. Clearly, if one knew that mechanism, one could write the
equations for it and those equations would have to be dimensionally valid,
but the dimensional transformation property allows the modeller to short-
circuit that difficulty.
MODEL MAPPING AND DOCUMENTATION
Explaining a model to oneself and to a client is difficult when working from

a listing such as Fig. 4.6. All the system dynamics packages therefore offer
some form of printout of the equations and the definitions. The DYNAMO
'standard' is shown below for part of the model in Fig. 4.6.
The dynamic equations, levels, rates and auxiliaries are laid out as shown
below , with the definitions of the variables and any constants required by
the equation. The constants do not have to follow the equation in the listing:
they will be searched for and printed out. This format is usually very easy to
explain even to a client who is not versed in computer modelling.
Equation number 1
r finra.kl=step(lOO,lO)+step(stepht,steptm)
c stepht=O
c steptm=120
FINRA=(TONSIDAY) FOOD INFLOW RATE

STEPHT=(TONS/DA Y) STEP HEIGHT IN FOOD SUPPLY
STEPTM=(DAY) TIME OF STEP INCREASE IN FOOD SUPPLY
Equation number 2
I food .k= food .j+ dt*(finra.jk - foutr .jk)

n food=O
c dt=O.25
FOOD=(TONS) STOCK OF FOOD

DT=(DAY) TIME STEP IN MODEL
FINRA=(TONSIDAY) FOOD INFLOW RATE
FOUTR=(TONSIDAY) FOOD OUTFLOW RATE
The complex input pattern 137
An additional form of documentation is the model map, a couple of lines of

which are shown below. The variables and constants are listed in alphabeti-
cal order, DT being the first alphabetically for this model. That is followed
°
by information on the line in the model" in which it occurs, the line in which
it is initialized (in this case, line means that DT does not require an initial
value on an N statement) and the line in which it is defined. After that, the
map shows the names of variables in whose equation it is used.
1 DT=(DAY) TIME STEP IN MODEL
Equation in line 22 Initialised in
Documented in 36
°
Used in equation for -
FOOD TIME
2 FINRA=(TONS/DAY) FOOD INFLOW RATE
Equation in line 6 Initialised in
Documented in 37
°
Used in equation for-
FOOD
The available packages differ enormously in the form in which this type of
information is presented. Some, for example, list all the levels, followed by
all the rates and then the auxiliaries.
THE COMPLEX INPUT PATTERN
Figure 4.11 used an extremely complicated pattern of food inflow to demon-

strate the idea of smoothing information.
The equation which produces that pattern is shown below. It uses several
of the standard functions and we shall use it to show how functions can be
combined to produce a desired result.
r finra.kl =step(base+amp*sin(6.283*(time.k- 30)/perd)*
x clip(1,O,time.k,30)*sample(noise(7),nperd,O),lO)
c nperd=2
c base=lOO
c amp=100
c perd=20
4 This book is about the principles of system dynam ics modelling and has, therefore, been
written to be as independent as possible of specific software. The line numbers given here are
correct for Fig. 4.6, but users of COSMIC will find that the line numbers printed out when
running this model are 2 larger than the numbers given here . This is due to the way that
particular package works and does not invalidate the principles explained here .
The STEP is used to generate the sudden onset of food supply at

TIME = 10. That which is stepped is everything from base to 0), at the end
of SAMPLE and the time at which the step happens is the 10 at the
end which has been printed in bold to draw attention to it. The
amp*sin(6.283(time.k-30) perd) is a normal sine wave, 6.283being the value
of 2*pi. However, in order to allow time for the smoothing of the STEP to
be seen, we wish it to commence at TIME=30 so, instead of just using
TIME.K, we use (TIME.K- 30), so that when TIME=30 the sine wave will
'think' that TIME=O and start the sine from O. Unfortunately, system
dynamics packages calculate every variable every DT so, from TIME=O to
TIME=30 the SIN function will be producing values, and those have to be
switched off, as we saw earlier. The STEP switches them off until TIME= 10
and the CLIP(1,0,TIME.K,30) does so between TIME=10 and TIME=30.
The effect of sample(noise(7),nperd,0) is to introduce random values
lying between -0.5 and +0.5. They multiply AMP so that AMP will be
between 50 and -50. These variations are added to the base level of 100, so
the effect is inflow rates lying between 50 and 150. The randomness is
sampled at intervals of 2 weeks to represent, perhaps, the frequency
of occurrence of good and bad weather on the route taken by the food
convoys.
IS THAT FUNCTION REALLY NECESSARY?
It is a great temptation to dress up a model by using lots of functions,

such as CLIP, TABLE, MAX and so on. Consider, however, Fig. 5,4,
which allows for the Food Ratio to range from 0 to 1. Suppose that
when the model is run, FRATIO never falls below 0.8. Clearly the data in
Fig. 5,4 and the equations which represent the curve are largely spurious
detail and the table should have been defined to be in the range from 0.7
(to be on the safe side) and 1.0, not from 0 to 1. Similarly, a CLIP is
supposed to offer a choice between two options . If it always chooses the
same one, then either the detail is spurious or, more likely, there is a
mistake in the model or the problem has not been correctly understood.
Careful study of output is essential both to remove unnecessary detail and
to detect errors.
SUMMARY
This chapter set out to be a brief introduction to the principles underlying

some of the techniques required in system dynamics modelling. The for-
mats used here are those which might broadly be regarded as traditional
in system dynamics, but they are not the only ones. At the risk of repeti-
tion, the actual details of using these principles will vary from package to
Summary 139
package and the user must consult the manual for the software being
used.
The reader was encouraged to test these examples in small models be-
cause they will be used, in some cases without further explanation, in the
models in the ensuing chapters.
CHAPTER 6
Case studies in modelling
INTRODUCTION
Chapter 2 explained the techniques and concepts of influence diagrams and

in Chapter 3 we studied influence diagrams for some complex problems.
Chapter 4 took us into the underlying theory of system dynamics modelling ,
including the building and running of a very simple model. Chapter 5 laid a
grounding in some of the essential techniques required to build models. In
this chapter, we must put those pieces of knowledge to work and build some
serious models of managed systems. At this stage, the emphasis will be on
building models and testing them to establish the confidence which can be
placed in them; the process sometimes called 'validation' , though 'confi-
dence building ' would be a far better term. The use of models in policy
design will be dealt with in Chapter 7 and their optimization in Chapter 9.
We shall start with the manufacturing company, the influence diagram
for which was studied in Chapter 2, because that model uses most of the
standard ideas of system dynamics and it proves to be very rich in possibil-
ities for policy design and optimization. That model will be developed in
small stages to show the procedure in detail , the process being applicable
regardless of the software package being used. Other models in this chapter
will be more in the nature of exercises for the reader.
THE DOMESTIC MANUFACTURING COMPANY (DMC)
The narrative for DMC in Fig. 2.4 dealt only with how the system worked,
because we were engaged in practising the techniques of influence dia-
gramming. Before we can build a model, however, we must be clear about
what it is for; that is, we must understand the problems faced by the
company's management. (This problem is based on one of the author's
consultancy assignments. It has been simplified, but the description of the
system and its problems is essentially accurate.)
The management problem

In this case, DMC's manufacturing activities were faced with a very unsta-
ble and completely unpredictable order pattern with two peaks and
troughs in each year . The company's historical experience is that, despite

this, the production department rode through the variations in the order
pattern quite well. The raw materials department, by contrast, could
barely cope . At times, there were serious shortages of materials, to the
point that the possibility arose of having to close down production. At
others, there was such a glut of materials that space had to be rented to
accommodate them . Naturally, this led to such interpersonal tensions
that the two managers concerned were scarcely on speaking terms. Apart
from the personal factors , one of the tests for the model will be to see if it
behaves in the same general way as the real system , and for the same
reasons.
The initial stages of modelling

Some initial steps need to be performed before one can start to write
equations. They are discussed in detail for this model, although readers will
have to apply them to the others in this chapter themselves.
The first stage is to choose names for the variables and to decide on their
dimensions . We shall use short mnemonic names, though some modellers
prefer long ones. Variable naming is a matter of personal preference, and
you should develop your own style, because a consistent style will help to
produce better quality models. The names and dimensions for this model
are shown in Fig. 6.1.
The second stage is to decide whether each factor in the diagram is a
level, a rate, an auxiliary , a constant or a table for a non-linearity. This
process is called type assignment and probably causes more problems for
the novice modeller than anything else, though it need not if the injunction
to think physics has been applied , the modeller is familiar with the idea of
level variables acting as memories in the model and has studied the com-
mon modules shown in Fig. 2.12.
To assign types, the place to start is the physical flows, shown with solid
lines in Fig. 6.1. Think of the analogy of water flowing and the places in
which it accumulates. Recall the use of terms such as 'inflow', 'outflow',
'drain' and 'level' in the common modules in Fig. 2.12. Thinking in that way
makes it evident that Backlog of Orders and Raw Material Stocks in Fig. 6.1
must be level variables and that the flows in and out of them must be rates.
Module B in Fig. 2.12 suggests that Raw Material Order Rate must also be
a rate variable. Bearing in mind module G, in which a rate and a parameter
affect a smoothed variable, the Average Order Rate and Average Produc-
tion Rate must also be levels, but of the smoothed variety which was shown
in Fig. 4.10.
Do not become confused by the word 'rate' in the names of the averages .
This does not mean that the variable is a rate. The key word is 'Average',
and averages in system dynamics are levels, because they are a memory for
what is averaged.
Avefaging Pefiod
for Order Rate
TAOR Averaging Period
(WEEK) _
tor Produclion Rale
'\ ,,/ <, TAPfl
'- (WEEK) ......
'\ /
/
-- ........ .. Average
Production - _
~AVerage \ Rate........ _ <,
Order APR
Rate \ (WMS/WEEK) .......... Desired <, .....
+ AOR + "" + Raw Material "'
'f" (WMS/wEEK) + r Stocks '" ........
/ <, .... Desired \ ~ DRMS "'
Backlog <, / r: (WMS) -, \
/ o tOrders I
DBLOG \- / // \
/ (WMS) I w..'" of Average \ \
Backlog / Production in • +
Discrepancy "- Desired Raw
/ » EPSI I / Malerial Stocks _ Raw Material \
/ +)"(WMS) W A P . . , Stock Disc repancy
\ (WEEK)./ EPS2 \
I Weeks of Average I /
Orders as Desired ./ (WMS)
\ Bocklog \
WAO / ,
I / - Raw Material
./ -, \. \
\ (WEEK) + / ,/ ~ Stocks \.
I +t' ~ / »> ~~~) -, I
\ I Production + -, I
Rate
\ PR +~ t'+
I -~(WMS/WEEI() Raw Material
\ + Backlog 01 4 Order Rate
Orders RMOR
BLOG I (WMS/WEEK)
(WMS) I ~ ~
Time 10 Correcl Raw Material \
Order Backlog A rrival Rate
Dilcrepancy RMAR \
TABL (WMS/WEEK)
(WEEK) +"" DDH Time10 Co"ecl
Raw Malefial
Dilcrepancy
TARMS
(WEEK)
Fig.6.1 Variable names and dim ensions for the DMC problem.
To emphasize those points , Fig. 6.2 shows the level variables with a
symbol denoting a bucket of water superimposed. Similarly, the rate vari-
ables have the engineering symbol for a valve superimposed', The param-
eters are shown in italics and the other variables must be auxiliaries . There
are no tables in this model.
Type assignment really is as simple as that. The main thing is to follow the
physics and identify the physical and smoothed levels, and the rest will
follow. Later in this chapter we shall state some rules for how variables can
affect one another, but the real key is to think physics and to be quite clear
about the principles embodied in Fig. 4.1.
Finally, one also has to identify any additional parameters which will be
needed and to choose whether or not to make the diagram more compli-
cated in order to show them explicitly. In this instance, extra parameters are
required to specify the pattern of New Order Inflow Rate, NO, but we
choose not to show these on the diagram as they will probably confuse the
client.
Some system dynamics software packages are based on drawing on the
screen, using icons, the level/rate type of diagram, a small example of which
was shown in Fig. 2.17. Using that type of software still requires the mod-
eller to make type assignments and also requires every detail to be shown on
the diagram. Whether this is helpful or laborious is a matter of personal
preference. The author's view is that models can be built just as effectively
and, perhaps, more quickly, with an influence diagram at about level 3 in
Fig. 2.16. The connections between the diagram and the model are clear
from the equations and the diagram is kept in its proper role as a tool for
communication with the client. These are, however, matters of opinion, and
readers must make up their own minds.
The common element in all the packages is that, as explained in Chapter
4, the equations can be written in any order the analyst finds convenient,
and the package will sort them into the correct computational sequence. In
practice, the modeller chooses an order which is easy to understand and
explain . Development of a consistent personal style will be a considerable
aid to speed and quality of work.
The model equations

We now turn to the writing of the model, explaining the equations as we go.
The principles are valid, though the details of entering the model into the
computer will depend on the package being used. As explained in Appen-
dix A, some packages will create some of the equations automatically,
though the user has to type the others or generate them using a mouse. For
the purposes of explanation, we shall proceed as though all the equations
I This symbology is not normal influence diagram procedure; it is used here only as an
explanatory device.
Averaging Period
for Order /late
TAO/l Averaging Period
(WEEK) for Produclion /lale
<, ,/ ....... TAPil
<, / (WEEK) <,
<,
- -- Average
Production
Rate
-------.....,
APR <,
(WMS/WEEI() ............ Desired <,
+ + Raw Material <,
/ 1,.....v, 'M~'k t Deslred Stocks " -,
/~DRMS
Backlog <, /
/ ot Orders / (WMS) "
DBLOG
/ (WMS) \- / Weeks of Aver~ \ \
Backlog Production in
Discrepancy .......
/ Desired /law
*l' + " \
/ » EPSl "' / Material Stocks Raw Material
/ +~(WMS) \ WAP - Stock Discrepancy \
I Weeks of Average / (WEEK) . / ;If' EPS2
Orders as Desired
/ ./ (WMS)
\ Backlog / ./ -, \
WAO / - ow Material '\
\ (WEEK) / Stocks \
I ,
--'-!...&I'--lI'''Y''-2
~.... ~
(WMS) '\ ' \
\ ~ I
I
\
\
Time 10 Correct
/law Material
Discrepancy
TA/lMS
(WEEK)
Fig. 6.2 Type assignment for the DMC problem.

are typed in by the analyst. In each case the set of equations is followed
by a brief explanation. Rather than repeating definitions each time a vari-
able is mentioned they are given in alphabetical order at the end of the
explanation of the equations. The reader will have to flip back and forth
between the equations and the definitions, though they are also shown in
Fig. 6.1.
* Fig 6.3 The Domestic Manufacturing Company Model

note
note file fig6-3.cos
The first few lines give a title to the model.
note
note current and average inflow of new orders
note
r no.kl=blev+d1 *step(20,10)+(1-d1)*amp*sin(6.283*
x (time.k-10)/perd)*clip(1,0,time.k,l0)
c d1=1
c blev=100
c perd=25
c amp=30
I aor.k=aor.j + (dt/taor)(no.jk-aor.j)
n aor=no
c taor=4
Next, we must decide how to drive the model with a pattern of new orders.
A STEP input is a standard driving force Which, as we shall see, gives useful
information about the model, but the company is known to have a pattern
of demand with two peaks and troughs in each year. We therefore also
provide for a sine wave and a 'switch', D1, to choose between them. If
D1 = 1, the form of NO will be a constant value of 100 up to TIME= 10, after
which the value will be 120. If D1 is set to 0, the behaviour of NO will be a
constant value of 100 until TIME=lO, after which there will be a sine wave
oscillating between 70 and 130.The equation for NO is rather similar to that
discussed on p. 137, except that it does not allow for randomness. At this
stage we do not wish to make the model behaviour too complicated to
understand; the ability to cope with random behaviour will be an aspect of
the system's robustness which the reader should examine in Chapter 7.
Note the extra parameters for BLEV, PERD and AMP , which are not
shown in Fig. 6.1. Their values are chosen to correspond to what happens in
the real system. Observe the use of NOTE statements as comments to
improve legibility and to identify the sectors of the model. As a rough rule
of thumb, about 20% of a model's lines should be comments.
The Average Order Rate is a standard smoothing equation, with the
smoothing done over 4 weeks. The )i( means multiplication and has the
146 Case studies in modelling
same effect as )*( . AOR is initialized to the external driving force so as to

set the model into balance before it is shocked by the STEP or the sine
inputs. Not all system dynamics packages will allow a variable to be used to
initialize a level in this way, but COSMIC will repeat the rate equation for
NO, so as to create the correct initial value for AOR. If your package will
not do this, the correct initial value is 100 [WMSIWEEK], as that is the
value of BLEV. That is, you will have to type in n aor=lOO.
note
note actual and desired order backlog
note
I blog.k= blog.j+ dt*(no.jk - pr.jk)
n blog=dblog
The influence diagram shows that the actual Backlog is fed by New Orders
and depleted by Production Rate. The positive and negative signs in the
influence diagram's portrayal of Backlog are mirrored exactly by the + and
- signs for NO and PRo It is a good method of modelling to check off the
links on the diagram as they are written into equations, making sure that the
correct signs are used.
The influence diagram also shows two influences on BLOG, and the
equation for BLOG has two terms within the ( ). Checks such as this,
elementary and obvious though they might seem, are essential to a good
model, and it is surprising how often they are not carried out.
Like all levels, BLOG requires an initial value. Rather than simply
setting it to be a number it is always worth considering what the system
is trying to do, where that can be worked out. In this instance, the produc-
tion manager tries to keep Backlog to its desired level, DBLOG. It is,
therefore, an interesting test to set Backlog initially to that target and see
how well the system manages to achieve its objective. As mentioned above,
COSMIC will repeat the auxiliary equation for DBLOG so as to initialize
BLOG. If your package will not accept this, the correct numerical value is
600 [WMS].
a dblog.k =wao *aor.k
c wao=6
The desired backlog is the Average Order Rate multiplied by the number
of weeks after which the customers will require their order to be
delivered.
note
note current and average production rates
note
a epsl.k=blog.k-dblog.k
r pr.kl=eps1.k/tabl+aor.k
c tabl=4
I apr.k=apr.j+(dt/tapr)(pr.jk-apr.j)
n apr=no
c tapr=4
The value of EPSl is fed into the Production Rate equation to give the
system dynamics format for the algebraic equation discussed on p. 25, with
the addition of AORK, as was explained in the narrative for this problem
in Fig. 2.4. The production manager's choice of policy is to correct the
backlog error, EPS1, over a period of 4 weeks.
It is essential to see that the choice of PR is a model of the production
manager setting the general level of activity in the factory, not of the
manager dealing with individual orders of specific types of washing
machines for particular customers. Although system dynamics can be used
to model the detail, it is more useful to take this broader view of
the information/action/consequences paradigm in order to get close to
the policies for running the business as opposed to the procedures of its
daily life. In short, one gives the production manager credit for know-
ing how do the job in detail and concentrates on the overall strategic
aspects.
Average Production Rate is Production Rate smoothed over 4 weeks.
The initial value for APR is set to the external driving force to make sure
that the model is properly balanced.
note
note actual and desired raw material stocks
note
I rms.k=rms.j+dt*(rmar.jk-pr.jk)
n rms=drms
a drms.k=wap*apr.k
c wap=8
note
note ordering of raw materials
note
a eps2.k=drms.k-rms.k
r rmor.kl = eps2.k/tarms + apr.k
c tarms=4
The raw materials manager 's sectors of the model are derived on exactly
the same principles as the production manager's, which were discussed
above. You are most strongly urged to cover up the raw materials
manager's equations given above and write them again for yourself using
the influence diagram and the production equations as a guide. No amount
of studying models will enable one to build them; the only way is to do it
oneself.
note
note delayed inflow of materials
note
r rmar.kl =delay3(rmor.jk,ddel)
c ddel=6
This uses a standard DELAY3 to represent the inflow of raw materials in
response to earlier decisions to order them .
note
note output and controls
note
c dt=0.25
c pltper=l
c prtper=lO
c length = 120
This section sets the parameters for controlling the simulation , the main
points being the choice of DT and of LENGTH. The former is calculated by
noticing that there are four first-order delay parameters, TAOR, TABL,
TAPR and TARMS , all of which have the value 4. The formula given on p.
109 would require DT=l for these parameters. However , DDEL features
in a third-order delay and has the value 6. That would require DT to be no
larger than 0.5, so we choose 0.25 to be on the safe side. LENGTH is chosen
to allow for just over 4 cycles from the sine wave, which should be enough
to show any characteristic behaviour. The parameters PRTPER and
PLTPER are the DYNAMO tradition for controlling printed and plotted
output, though different methods are used in other packages.
print 1)no,aor
print 2)pr,apr
print 3)blog,dblog,eps1
print 4)drms,rms,eps2
print 5)rmor,rmar
plot no=a,pr= b,rmor=c(0,250)
plot blog=a,dblog= b,rms= c,drms=d(0,2000)
The first five of these statements produce printed tables of all the model
variables. There is absolutely no substitute for printing out tables of the
variables, and actually looking at the results, if one is to debug a model
properly.
The last two lines plot graphs of a selection of variables, using the COS-
MIC techniques for flat-bed plotter output in order to produce the graphs
which will be seen in Fig. 6.3.
note
note simulation experiments
note
250
225
200
175
150
125
100
75
50
25
1 1~2
TIME (WEEK) SIMLl.ATED TIME
(WMS/WEEKI NEW ORDER RATE
(WMS/ WEEKI PROOUCTION RATE
- ·RMOR (_/WEEKI RAW MATERIAL ORDER RATE
2000
1750
~o . 90. ~o. 50. so. 70. Bo . go . 'lOa . ~ 10 . ~ 20

T I ME (WEEK) SIMLl.ATED TIME
(_ ) ORDER BACKLOG
(WMS) OESIIlEO BACKLOG
I_ I RAW MATERIAL STOCKS
I_I OESIIlEO RAW MATERIAL STOCKS
A - STEP RESPONSE OF MODEL
Fig. 6.3 Th e dom estic manu facturing compa ny.

250
225
200
. '10. 20.
I:L TI ME (MEEK) SIMlLATED TIME
IL RMOR
~.
R
NpO (lIMS{WEEK) NEll - RATE
(lIMS{WEEK) PRODUCTION RATE
(lIMS{WEEK) RAM MATERIAL CAllER RATE
2000
. . -M:"-. ~o . ~o . 50. 130. 70. 130 . bo o 1100. 1110. ~20 .

II :L TIME (WEEKI 9IMIUTED TIME
1
'. 1 : BLOG
...::...: 2a~OG
(WMS) OAlJER 1lACKL06
(MIlS) DESIRED 1IAC1<LD6
1-. _ DRMS l::l ::I:::~ERJ~~=AL 9TDCKS
B - RESPONSE TO SINE WAVE
Fig. 6.3 Continu ed

:::~:
200 I
I
175 I
150 Ii,
125
100 -t;-t--~
75
50
25
TI ME (WEEK) SIMULATEO TIME

(WNS/WEEK) NEW DIllER RATE
(WNS/ WEEK) PROOUCTIOM RATE
(WNS/WEEK) RAW MATERIAL ORDER RATE
2000
1750
. 20 .
TI ME (WEEK) SIIU.ATEO TIME
(WM91 ORDER BACKLOG

(WNS) DESIREO BACKLOG
(WWS) RAW MATERIAl. STOCKS
(WMS) DESIREO RAW MATERIAl. STOCKS
C - LOWER STOCKS AND MORE SMOOTHING

Fig. 6.3 Continu ed
run A - Step Response of Model

c dl=O
run B - Response to Sine Wave
c wap=4
c tarms=8
c tapr=8
run C - Lower Stocks and More Smoothing
Finally, we have three experiments (which could be performed interac-

tively) which will show first the STEP response, secondly the reaction to a
sine wave and, thirdly a 'what would happen if' experiment in which the
company tries to hold lower stocks of four weeks average production and
two of the policy reactions are slower.
note
note variable definitions
note
d amp=(wms/week) amplitude of sine wave in new orders
d aor=(wms/week) average order rate
d apr=(wms/week) average production rate
d blev=(wms/week) base level for order patterns
d blog=(wms) order backlog
d dblog=(wms) desired backlog
d dl=(l) demand pattern switch
d ddel=(week) delivery delay on materials
d drms=(wms) desired raw material stocks
d dt=(week) solution interval
d epsl=(wms) difference between backlog and desired backlog
d eps2=(wms) difference between desired raw material stock and
x actual raw material stock
d length = (week) simulated duration
d no=(wms/week) new order rate
d perd=(week) period of sine waves on new order pattern
d pltper=(week) plotting interval
d pr=(wms/week) production rate
d prtper=(week) printing interval
d rmar=(wms/week) raw material arrival rate
d rmor=(wms/week) raw material order rate
d rms=(wms) raw material stocks
d tabl=(week) backlog adjustment time
d taor=(week) order averaging time
d tapr=(week) production rate averaging time
d tarms=(week) adjustment time on raw material stocks
d time = (week) simulated time
d wao=(week) weeks of average orders in desired backlog
d wap=(week) weeks of average production required in raw material

x stocks
The actual format will depend on the package , but the principle of writing
definitions that are as clear as possible for every variable and parameter is
a sound one. A model which contains undefined variables is useless, and
one which contains variables simply called 'X' is not much better. Some
modellers prefer to put the definitions immediately after the first use of the
variable to which they refer, but the author's practice is to group them in
alphabetical order at the end of the model , especially with a model of any
appreciable size.
Model validation
So much for developing the model's equations. But how do we know they
are right ? A number of tests can be applied and , the more tests are passed ,
the more one's confidence increases that the model is technically correct in
terms of the rules of system dynamics equations, and that it corresponds to
the real system well enough to be used to make policy recommendations to
its managers.
On pp. 96-8 we suggested some tests of confidence in a model, and this
issue is so important that it is worth restating them so that the reader can
apply them to this problem.
• The influence diagram should correspond to the statement of the prob-
lem. We have already established that it does .
• The equations should correspond to the influence diagram. The reader
must check that they do. Part of this test is, of course , to ensure that there
are no errors of syntax, such as missing parentheses, variables which are
used without having been defined or variables which are defined or
documented without being used. The reader will have to run the model
to verify that.
• The model must be checked for dimensional validity, either by hand or
by software .
• The tests for mass balances were deliberately not included in this model
to encourage you to write the extra equations yourself, not neglecting the
amount of raw material in the delay contents between RMOR and
RMAR.
• The tests that the model should not produce any ridiculous values and
that its behaviour should be reasonable are seen in the three pages of Fig.
6.3.
Figure 6.3A2 shows the step response, and all the plotted variables seem
reasonable. That is not a sufficient test, as not all the variables are plotted,
2 Recall the explanation on p. 95 for the conventions for reading these graphs. The style of
graphical output varies between packages; this is COSMIC's.
and in a model even a little larger than this it would be impractical to do so.
One has, therefore, to look at the printed tables . None of the printed values
is implausible. However , intuitively, one would not expect to see negative
values in this physical system, but EPSI and EPS2 do sometimes go nega-
tive. Why is that acceptable?
The dynamics of the step response cannot be compared with the
real system as it never experiences such a simple shock. Even so, the
dynamics show some important results which we have to treat as interim
results until we have tested the model against a more realistic pattern. The
first is the extreme sluggishness of the response . Even by TIME=lOO,
nearly two years after the step, the oscillations are only just dying away. The
second is that the production variables, PR and BLOG, respond reasonably
smoothly to the step and BLOG and DBLOG match very closely. This is
consistent with the company's experience that the production department
copes reasonably well with external shocks. On the other hand, the raw
materials variables show a very pronounced reaction to the step, consistent
with the experience that they do not cope very well. Furthermore, the time
from one peak of the oscillations in RMOR and RMS to the next is about
25 weeks. This is known as the natural period of a system. In this case, the
natural period is the same as the period of the cycles in demand, and it is
usually a recipe for extreme instability to drive a system at its natural
period. The film of the breakup of the Tacoma Narrows bridge is a dramatic
example of that phenomenon.
The sine wave response in Fig. 6.3B is even more worrying, but, paradoxi-
cally, also more convincing.
The worry is that the response of RMOR in the top half of the graph is
explosive, to the point where , at TIME=32, the graph goes off the page,
thereafter going further and further apart. Similarly, the graph of RMS also
shows the explosive oscillation, which was predictable from the natural
period deduced from the STEP input results. RMS and DRMS only match
eight times in two years, and are completely out of phase . For the produc-
tion department, on the other hand, the graph of PR runs fairly smoothly
through the middle of the sine wave of NO, and BLOG and DBLOG match
very closely.
The paradoxical conviction is that the model is broadly on the right lines
because this is roughly how the real system behaves. The production people
can cope with the demand variations, but those in raw materials have a very
hard time, which is what the model suggests. On the other hand , the explo-
sive behaviour will eventually lead to such outrageous consequences as
negative raw material stocks, so we are faced with the problem that, though
the model may be on the right lines, it is clearly unrealistic because a real
system cannot produce negative stocks.
This dilemma means that, on the face of it, we ought to make the model
much more complicated to prevent these unrealistic phenomena from aris-
ing. On the other hand, we could adopt the viewpoint that the unrealistic
behaviour arises because of poorly designed policies in the system and that
the measures which the people in the real system would have to take to
prevent negative stocks arising would be no more than desperate attempts
to cope with catastrophe. It therefore seems better to spend our efforts in
designing the catastrophic behaviour out of the system rather than trying to
model it.
This is a fairly fundamental point about the attitude to 'validation' . A
valid model is one which is, as was stated on p. 12,
well suited to its purpose and soundly constructed
We have satisfied ourselves that the model is technically sound by making
sure that its equations agree with the influence diagram, that they are
dimensionally valid and that the model's behaviour is, in important re-
spects, 'like' that of the real system. Its suitability for its purpose is largely
a matter of its simplicity, clarity, acceptability to the client and whether it
contains the policy options which will lead to improvement. In other words,
does the model contain the solution to the problem?
Testing the model

We can easily test this by the 'what would happen if' experiment which was
set up within the model and which is shown in Fig. 6.3C. This shows a
system which is well under control and suggests that the explosive behav-
iour arises from the exaggerated response which the original policies made
to the unpredictable external driving force.
Although Fig. 6.3C is a considerable improvement over Fig. 6.3B, the
behaviour it shows is very far from perfect. The oscillations in Raw Material
Order Rate are as large as those in New Orders and much larger than those
in Production Rate. In the lower half of the graph , the match between
BLOG and DBLOG is less close than in Fig. 6.3B, and actual and desired
Raw Material Stock still match only four times a year. RMS is worryingly
small at some stages of the run. There is scope for improvement, which we
shall study in Chapter 7, but the model does contain solutions to the
problem of bringing this system under control in the face of unpredictable
external demand.
For the present, we shall leave the Domestic Manufacturing Company,
which is on the model disk as FIG6-3.COS, though we shall use it again in
Chapter 7 for policy design and in Chapter 9 for optimization.
THE MANAGEMENT CONSULTANT'S PROBLEM
The influence diagram for the consultancy company's problem discussed in

Chapter 3 is reproduced for convenience in Fig. 6.4. The problem requires
-- - -- - - -- -- -- - - -
-------- <,
<,
"- <,
<,
"-
"- , -,
Ba<",
Demalld tor \
Consultancy
S9N1C es
\
\
I
I
- ...:. ENOf1_ ..... / -_ SLIGndafa I
1«000 - - - - - I'foIecI
,
I
,
I
, , N_oI
I
- _ " Guc:IIIId I
--
-- - -_ -_ _ L..~
' 1I«l"- ".~
~
--- /
Quol tl8d ... ...../
- - - - _ _ CorliUf o nts -
I'J8f ProJOCt
~+
and~ I
/
==
/
/
_ /
/
/
\ /
/
/
Fig.6.4 The consulting firm's problem.

some new modelling techniques and has some very interesting behaviour.
Again, the equations are shown in groups, each of which is followed by
an explanation. The reader should study the group before reading the
explanation.
The management problem
What, however, is the problem? The narrative on p. 49 mentioned that the

firm felt that it was falling behind its market potential and that it never
seemed to have the right numbers of trainees and consultants. This is rather
vague, but it is typical of practical problems in that the problem proprietors
often find it hard to articulate their difficulties. It sometimes turns out that
they are worrying unnecessarily. We shall see what happens when we run
the model.
The model equations

The reader should first study Fig. 6.4 and decide on variable names, work
out the dimensions of each variable and assign types to the variables. The
variable names used below are mnemonics based on the initial letters
of the names used in the diagram. There are four levels and six rates, but
it is vital that you try to work these out for yourself before looking at
the solution . Nothing is learned by studying someone else's solutions; you
have to make your own attempt and work out where, if at all, you were
wrong.
The equations are discussed below, with the definitions given at the end
of the explanation. The model name is FIG6.6.COS because Fig. 6.5 will be
introduced before the model's graphs are shown.
* Figure 6.6 The Management Consultant Problem
note
note file named fig6-6.cos
note
note
note demand and business generation sector
note = = = = = = = = = = = = = = = = = = = =
note
note basic demand
note
a bdc.k= base + swdp*(ramp( slp,slptim) + amp*
x sin(6.283*time.k/bcylen»
c base=50
c amp=lO
c swdp=O
c slp=O.l
c bcylen=48
c slptim=10
After the usual opening lines note the way in which the model has com-
ments for its major sectors and subsectors to improve clarity of explanation.
BOC is modelled to offer a choice between constant demand and a rising
oscillatory trend. The former is artificial, but is useful for debugging and
testing, as the behaviour is not clouded by the sine wave effects.
It is always important to be aware of any hidden assumptions in a set of
equations. In this case it is that BOC represents the demand for this firm's
services; its share of the total market. As we shall see below, the firm may
get more or less than its normal share , depending on the service it offers to
clients, so the actual Rate of Obtaining New Projects may be greater or less
than BOC.
Most consulting firms operate on a basis of charging for the time spent on
a project at a rate which covers all costs and a profit margin. The fact that
BOC exists suggests that clients are content with this arrangement, so
money is, if not irrelevant, at least peripheral to this problem.
note
note effect of qualified consultant availability
note on rate of obtaining new projects
note
a qcpp.k=qcafp.k/nproj .k
a cpprat.k=qcpp.k/nnqcpp
c nnqcpp=1.82
These variables and equations are simply those in the upper right-hand
corner of Fig. 6.4. The sector is broken into several auxiliaries to make
writing and debugging easier. Because the names state pretty clearly what
the variables mean , the equation writing can flow easily. For example,
Qualified Consultants per Project can only be the number of qualified
consultants not supervising trainees divided by the number of projects.
Careful attention to variable names will considerably assist modelling,
partly for the reasons just given but also because clear names make it easier
to write the correct dimensions and to get the confidence which comes from
the model being known to be dimensionally valid, because the dimensional
validity can more easily be checked by hand or by software.
a cpmult.k=tabhl (ctab, cpprat.k, 0, 2, 0.5)
t ctab=0/0.3/1I1.5/1.7
These equations represent the functioning of the link from the Consultants
per Project Ratio to Rate of Obtaining New Projects in the top right-hand
corner of Fig. 6.4.The rise and fall of CPPRAT represents the success of the
firm in providing the right number of qualified staff to supervise the projects
on hand , and this, as the problem description makes clear, affects the firm's
ability to generate new business. As will be seen below, the multiplier ,
CPMULT, operates on the Basic Demand for Consultancy Services to help
the firm to win or lose market share . This new variable, CPPMULT is not
shown on the influence diagram and this is a typical case of drawing a
diagram at something less than Level 4 so as to convey the right information
to the client without confusing the issue with extraneous detail. For prac-
tice, the reader should alter Fig. 6.4 to show this new variable, the solution
being shown in Fig. 6.5.
The TABHL function is used in preference to TABLE as there is no
reason why CPPRAT should not exceed 2. The numerical values in the non-
linear table CTAB are found by studying previous sales data and relating
them to data on numbers of staff available, by asking experienced managers
their views and by common sense. It is, for example, clear that no work will
be obtained if there are no consultants, accounting for the 0 in the first
element of CTAB. By definition, when CPPRAT= 1 the situation is normal,
so the third entry in CTAB is 1.
The important point is that the influence diagram's portrayal of the
mechanisms of obtaining business has been agreed by the company as
reasonable and it now forms a theory on which data analysis can be based.
It is not a matter of collecting some data and seeing what theories are
suggested; it is a question of having a theory and analysing data with its
assistance.
Note also that NNQCPP is correctly defined in the D statements below as
[PEOPLE/PROJECTS]. It is very common to see numbers carelessly called
'ratio' and given dimensions of [1] and that is usually a sign that the model's
dimensions have not been properly checked . The ratio CPPRAT is, how-
ever, correctly dimensioned as [1].
It would also be easy to make the typing error
a cpprat.k=qcpp.k*nnqcpp
which is clearly wrong, because CPPRAT should be a ratio but [PEOPLE/
PROJECTS]*[PEOPLE/PROJECTS] is not [1]. Such errors can be found
by dimensional analysis, but only if dimensional checking is actually
done.
note
note projects sector
note ========
note
note projects being obtained and started
note
r ronp.kl = bdc.k*cpmult.k
r npsr.kl =delay3(ronp.jk,cdel)
c cdel=2
--- -- -- -- - - - - - --- --------------
/ " --
-- ... ... -,
/
/ ... ....
/
I "
I Bcsc
De mond for
I
" ,,
Con suttoncy \
I ServICes
t+ \
Re QJife d I
Number o f I
,.. - Consunont ,
/ e"'JlVoienh '" ... Slondan:l oIC_
_IorEnect
_. _
/ ~ - - - - - - - f'fof«:/ I
/ o nn. _ _ - Profect on llol. 01 ,
I
... ...
- 0bIaInJng N.w
f'fof«:" I
I _ -- - _ I
I " - I
+ + .:- I
Manpow e r -- /
Discre panc y
Q uohf le d "" _ /
\ C on sultcr. f$ -
\ per Proje ct
\ PlOQIctlv/ly /
/ 1+
\
\
":::::'-. /
"
.
I
/
\ /
nn./o \ / -- "
carr.,,/ / ---
O,"",-y \ / ---
\ /
\
\ + /
" ... /
-. DeSired ,.
,....... +
...Recrultment
Rate
.
/ I
/ I
I \
I \
Ave~oge
Recruitment
-*' \
Ra1e \
.... + ,
\ -: -.
I '" ... -, ...
"\ ------ D EMPlO Y + II c~;~:~~nt
Rate
/leCruII .... nI
Avetaghg
P.rIod
Fig.6.5 Th e co nsulting firm's problem (modified).

The projects sector is a straightforward flow of projects starting after a delay
for contracts.
I netdef.k=netdef.j +dt*(bdc.j -ronp.jk)

n netdef=O
A new variable is introduced to measure the cumulative difference between

projects available, BDC, and those obtained, RONP, as a measure of the
extent to which the firm is, or is not, falling behind its market potential. A
positive value of NETDEF means that the company is not meeting all its
demand, while a negative value would mean that market share is being
created. NETDEF is not shown on the influence diagram as it is not part of
the system. It is, in fact, an additional variable introduced to help the
analyst to gauge system performance.
Note that this involves feeding an auxiliary variable, BDC, into a level,
whereas so far we have seen only rates feeding levels and, indeed, Fig. 4.1
implies that auxiliaries cannot feed levels. This is an example of how the
flexibility of most of the system dynamics packages allows one to bend the
strict rules to simplify modelling provided one understands the rules and
knows when and why they can be broken. Using the flexibility of the soft-
ware to fiddle one's way out of a difficulty is very bad practice. In this case,
the alternative would seem to be to define BDC as a rate , but that involves
the difficulty that rates must be reserved for things which actually flow into
or out of physical levels in the real system. NETDEF is not such a level; it
is an artefact created by the modeller to help measure system performance.
In technical terms it is a simple example of the objective functions which
were mentioned in Figs. 1.3 and 1.4 and which we shall encounter in more
detail in Chapters 8 and 9.
In the equation for NETDEF a typing error such as
I netdef.k = netdef.j + dt*(bdc.j+ronp.jk)
could not be detected by dimensional analysis, but only by running the

model and carefully examining why NETDEF always turned out to be very
strongly positive. Dimensional analysis is a necessary but not sufficient
condition that a model is correct, and there is no substitute for carefully
studying the output and asking oneself whether this is the same thing that
the real system would have done, and for the same reasons.
note
note projects in progress
note
I nproj.k=nproj.j+dt*(npsr.jk-pcr.jk)
n nproj=242
A simple level for projects. The initial value of 242 is derived from company
data, as are the other numerical values.
note
note project completions
note
r pcr.kl =nproj .k/dproj .k
a dproj.k=spt*effrat.k
a effrat.k=recpp/acepp.k
a acepp .k=ecw.k/nproj.k
c spt=5
c potas=O.5
c recpp=2
A deceptively simple sector in which DpROJECf is modelled as dependent on

the variables identified in the influence diagram, though the numerical
value for RECPP was not mentioned in the narrative. The reader should
work through the equations from DPROJ to ACEPP, checking against the
influence diagram and understanding how they are derived. The definitions
and dimensions appear below. After working through them they will look
like very simple common sense and we have already remarked that disci-
plined common sense is very helpful in doing system dynamics. Most of the
technique is very simple, because unless we can work in simple steps we
shall have little practical chance of modelling the complexities of managed
systems.
The key feature of this sector is that PCR is modelled as a first-order
delay of PSR , as can be seen by comparing the equations for NPROJ and
PCR with Fig. 4.8(a). It is important to remember that the D on a delayed
link in an influence diagram does not state what type the delay is, it merely
states that there is a delay. Why, though, did we use DELAY3 for the
connection between RONP and PSR while here we have not even used
DELA Yl but have explicitly written the equations for a first-order delay?
There are two reasons .
The first is that DELAY3 is standard in system dynamics and requires
only one equation. The second is that we do not need the contents of the
delay between RONP and PSR, but we do need the contents between PSR
and PCR, because they drive several other factors in the model. Modelling
the contents, NPROJ, requires an initial value , and that value would have to
be exactly consistent with the three internal levels of DELAY3, had we
used that. Getting initial conditions to balance in such equations is very
difficult, so it is easier to do it in an explicit first-order delay.
The general rule is that, where the contents of a delay affect its input or
its output, the delay equations have to be written explicitly in order to get
the initial conditions to balance correctly.
Some packages will allow the contents of delay functions to be available,
but it is better to make things explicit. In any case, as we saw in Fig.
5.2, there is not very much difference between the behaviours of the
different delay types when there is a continuous input to the delay, which is
why we did not add complication to the model by writing the explicit
equations for a third-order delay. System dynamics concerns itself with
overall behaviour under the influence of policies, not with an illusory
attempt to get very precise values when the data are usually fairly imprecise
in any case.
note
note trainee consultants sector
note ==============
note
note desired trainee recruitment rate
note
a ecw.k=qcafp.k+potas*(tcbt.k+qcst.k)
a dtcrr.k = (mandis.k/potas)/rat + arr.k
a mandis.k=rnce.k-ecw.k
a rnce.k=nproj.k*recpp
c rat=lO
A set of auxiliaries corresponding closely to the influence diagram. Again,
the reader should work through them.
note
note non-negativity constraint on recruitment
note
r trr.kl=max(dtcrr.k,O)
note
note average recruitment rate
note
I arr.k=arr.j+(dt/rat)(trr.jk-arr.j)
n arr=4.0
The non-negativity constraint is modelled as was explained on p. 123. The
reader should review that section to recall the difference between the two
types of non-negativity constraint. There is a simple smoothing equation for
ARR.
note
note trainee consultants being trained
note
I tcbt.k=tcbt.j +dt*(trr.jk-tcr.jk)
n tcbt=96
A simple level for TCBT .
note
note trainees becoming qualified
note
r tcr.kl=tcbt.k/dtrain
r trqs .kl=tcr.kl*psat
c dtrain=24
c psat=0.9
TCR is modelled as a first-order delay, for the reasons explained above .
The equation for TRQS is another example of bending the rules of Fig. 4.1,
in this case to allow a rate to affect a rate. TCR is a physical flow from
TCBT and TRQS is also a physical flow into QCAPT. Multiplying by PSAT
simply recognizes that one is not the same as the other.
note
note qualified consultants sector
note ===============
note
note number of people
note
I qcapt.k=qcapt.j+dt*(trqs.jk-clr.jk)
n qcapt=441
note
note consultant leaving rate
note
r clr.kl=qcapt.k/demploy
c demploy= 120
This group of equations is easily followed by analogy with the trainees. In
effect, the trainees' sector and the consultants' are identical modules , and
recognizing such cases is a useful skill as it speeds up modelling, provided
one is very careful not to jump to conclusions and too readily assume that
modules are identical without thinking about it. Some packages support
vectors and arrays , though these features should be used only after careful
thought about the problem.
note
note consultant division of labour
note
a qcst.k=tcbt.k*ratstt
c ratstt=O.l
a qcafp.k=qcapt.k-qcst.k
This is simple arithmetic to subtract from QCAPT the small number of
consultants who are supervisors. With 96 trainees, about 10 consultants will
be required for supervision, which is so few out of 440 consultants that this
aspect of the narrative is probably irrelevant and might, on the face of it, be
dropped from the model. The reason why it cannot be dropped is that the
partner in charge of training is a very powerful and senior figure in the firm.
It is highly unlikely that he would take the slightest notice of the model if his
interests were not represented in it. Such political aspects often playa role
in practical modelling. Indeed, it is sometimes a fine judgement to decide
between including detail for political reasons and making the model so
complicated that it ceases to be well suited to its purpose of policy evalu-
ation and design.
note
note control and output
note = = = = = = = = = = =
note
c dt=0.125
c prtper=5
c pltper=l
c length = 120
print 1)bdc,ronp,netdef,npsr,pcr
print 2)nproj ,qcpp,cppra t,cpmult,dproj
print 3)qcapt,qcafp,qcst,clr,acepp
print 4)dtcrr,trr,arr ,tcbt,trqs
print 5)effrat,ecw,rnce,mandis
plot bdc=a(0,120)/nproj = b(0,400)/qcapt=c,
x tcbt=d(0,600)/netdef=e(0,300)
DT is governed by the shortest delay in the model, CDEL, which is 2
months long and third-order. That suggests that DT cannot be greater than
116, to which 0.125 is the next lower acceptable value. The value of
LENGTH is set to the employment period of consultants, which is the
longest delay in the model , and should allow all the interesting dynamics to
be seen. (Experiment with LENGTH=240.) All the variables are printed,
for debugging, and a selection are plotted. The commands used in the
PLOT command are special to COSMIC, and your package may be differ-
ent. The scale for NETDEF is suitable for the two cases tested here, but, in
general, we have no idea what its magnitude will be, or even whether it will
be positive or negative.
note
note
run A Constant Demand
c swdp=l
run B Demand Growth and Business Cycle
Two experiments to create Fig. 6.6.
note
note definitions of variable s
note
d acepp=(people/projects) actual consultant equivalents per project

d amp=(projects/month) amplitude in the business cycle
d arr=(people/month) average recruitment rate of trainees
d base=(projects/month) base level of demand for company's services
d bcylen=(month) length of the business cycle
d bdc=(projects/month) basic demand for consultancy
d cdel=(month) delay in signing consultancy contract
d clr=(people/month) consultant leaving rate
d cpmult=(1) multiplier reflecting the effect of consultant availability
x on the rate at which new projects are obtained
d cpprat=(l) ratio of ability of consultants to do projects to number
x of projects on hand
d ctab=(l) table of values for cpmult
d demploy=(month) consultants' average employment time before
x leaving the company
d dproj = (month) actual project completion time
d dt=(month) solution interval
d dtcrr=(people/month) desired trainee consultant recruitment rate
d dtrain=(month) training delay
d ecw= (people) equivalent consultant workforce
d effrat=(l) ratio of required and actual equivalent consultants per
x project
d length = (month) simulated duration
d mandis=(people) manpower discrepancy
d netdef=(projects) net deficit between BDC and RONP
d nnqcpp=(people/projects) normal number of qualified consultants
x per project
d nproj=(projects) current number of projects
d npsr=(projects/month) new project start rate
d potas=(l) productivity of trainees and supervisors
d pcr=(projects/month) project completion rate
d pltper=(month) plotting interval
d prtper=(month) printing interval
d psat=(l) proportion of trainees who stay with the company on
x completion of training
d qcafp=(people) qualified consultants available for work on projects
d qcapt=(people) number of qualified consultants available for
x projects or training
d qcpp=(people/projects) qualified consultants per project
d qcst=(people) qualified consultants supervising trainees
d rat=(month) recruitment adjustment time
d ratstt = (1) ratio of consultants to trainees
d recpp=(people/projects) required equivalent consultants per
x project for normal project duration
d rnce=(people) required number of equivalent consultants

d ronp=(projects/month) rate of obtaining new projects
d slp=(projects/month) slope for growth in market
d slptim=(month) time at which slope is allowed to start
d spt=(month) standard project time
d swdp=(1) switch to turn demand pattern on and off
d tcbt=(people) trainee consultants being trained
d tcr=(people/month) rate of completion of training
d trqs=(people/month) transition rate of trainees to qualified status
d time=(month) simulated time
d trr=(people/month) trainee recruitment rate
As with the previous model , the definitions are in alphabetical order at the
end of the model.
Model validation and testing

The behaviour of the model with constant demand and with a rising sine
wave is shown in Fig. 6.6. With constant demand (Fig. 6.6A), the company
declines very slowly in terms of numbers of consultants and trainees. The
number of projects not obtained, NETDEF, reaches about 200 over 10
years, though that is a small proportion of the 6000 projects which might
have been obtained. On the whole, the model is broadly consistent with
the reference mode of the behaviour of the firm as described in the
narrative.
When the sine wave is implemented (Fig. 6.6B), to simulate a slowly
rising market with an oscillation due to the business cycle, the behaviour is
more complicated. The number of consultants, QCAPT, oscillates gently,
though the peaks and troughs are almost exactly out of phase with those of
BDC. Since there are relatively large numbers of consultants during the
troughs of the cycle, there is a surge of new projects at the beginning of the
upswing of the cycle, for example, from month 40 to month 50. As the cycle
reaches a peak around TIME=60, the Rate of Obtaining New Projects
ceases to rise sharply and there is a slow decline in the rate of obtaining
work until the next upswing of the cycle. The numbers of trainees rises
sharply just as the cycle passes its peak . NETDEF, the projects lost, oscil-
lates quite severely, with numbers of projects lost during the peak and
downswing of the cycle (from TIME=O to TIME=25, for example). In the
early stages of the upswing, market share is built and NETDEF falls to a
minimum at about TIME=48. At the end of 10 years, NETDEF is 289,
though that can only be found from the printout.
Overall, the simulation corresponds to the rather vague worries that the
company does not have the right numbers of people at the right times. What
can be done?
It is quite obvious that the company needs to find some way of increasing
400 800 500 15 0
l 11 1
35017oo~ l 450 125
300~ II l 600 400
250~ 11!I 350 ~

: l 300 ~ 500 I 100
200 ~ 400~ 250~ 75

- - - - - - - - - - - - - - - -- - - - - - - - - - - - - - - - ~ .~.~.~~~
150J 3001111 200~
I 200~
50
150~
I
100..; I
100~ 50~l
'I 100 25
50 i
, Jjl I
oJ 0 ' oJ 0
. 20 .
I IIII !
I BOC
L' TIME lNONTH) SIIO..UTEO TIME
lPRO.JECTS/I4OHTH) aAsIc DEMAND FOR CONSlA,TANCY
------ NPROJ (PROJECTS) ClRlENT NlHER OF _JECTS
, ~ - - -- QCAPT (P£OP\.E) NlHER OF OUAclFlEO CONSUl.TANTS AVAICAl!CE
"L"-==:'===--~~~¥6EF ~~gsr~~gE~~~~~~E~~ ~~I:'

A - CONSTANT DEMAND
400 800 500 150
l 11 1
350 1700~ 450 l 125
300J 600 II 400 l

! ~11 350~ /
100
"-
'250 ~ 500 '
:
200~ 4oo~ 250~
I
300.,:
75
--- --------~-----
_.........------.... ,,"
-----------L
------ --.. . . . " /
,'-- . ---.
.....
,, / /
150 J 300 '1 200~

.. ... ..
'-~
...,,"
. \.
.r ··.. . . . . . . . --- :
I 2oo~11 1!50~
50
/
100..;
I '\ 100~
I
25
r>::
./ "\
/ .._. . . . . -,j
/' . .
_ .
50i 100~ 50J

./ ......... \. <, <, /
. JJI I
/ -.....~.--./-1 »<:»
oJ oJ o.l 0
o>~~. to. 70 . ~o. 90. 'lOa . 1110 . ~20.
II
I'I :
:
BOC
L' TI ME (MONTH' SIIO..U TEO TIME
(PROJECTS/ NONTHl aASIC DEMAND FOR CONSlA.YANCY
.~
~._
-._
.==-= mB¥ (PROJECTS) ClSlRENT Nl.NllER OF PROJECTS
(PEoPLE) Pf..IiI8EA OF QUALIFIED CONSU...TAHTS AVAILABLE
'--.._ .._ .._ .._ .NETOEF
• .- .- TCBT (PEoPLE) TRAINEE CCHSULTANT9 BEING TRAI NED
(PROJECTS' NET DEFiCIT BETWEEN 80C AND RONP
B - DEMAND GROWTH AND BUSINESS CYCLE
Fig. 6.6 The management consultant problem.
the productivity of trainees and consultants, and the effects of doing so are
shown in Fig. 6.7A in which POTAS=O.7 . The striking thing is that
NETDEF increases, the value at the end of 10 years actually being 649! The
company is doing worse with more productive trainees. One of the standard
system dynamics tests is to subject the model to extreme, even ridiculous,
values to see if its behaviour is correct for those conditions. Figure 6.7B
shows the effects of lower productivity (POTAS=O.3) of consultants and
trainees, the result being better in that NETDEF becomes negative , reach-
ing - 379 by the end of the run. A negative deficit means that the company
is building market share. It is contrary to intuition that improving produc-
tivity makes matters worse and that reducing productivity makes things
better, hence the caption 'Counter-intuitive Consultants' on Fig. 6.7.
A further insight into this strange behaviour is given by the second part
of Fig. 6.73 on p. 171 which shows comparative plots for the total number of
qualified consultants, QCAPT, and for the net deficit, NETDEF, against
the three cases of productivity. The salient feature is that lower productivity
leads to better market share and to more consultants, whereas higher
productivity does the opposite. A little thought shows the reasons . With
lower productivity, more trainees are recruited who, after about two years,
become qualified. Since it is qualified consultants not supervising trainees
who generate business, it is not surprising that more business is generated.
The reason for the effect is the non-linearity in CTAB, and we have no
really solid data on CTAB. One might worry, therefore, that the results are
not 'correct', since CTAB is not known accurately. While it is certainly true
that changing the values of CTAB would change the detailed results, it
would not change the type of result as long as the shape of CTAB is correct.
No-one in their right mind is going to set out to reduce productivity; the
implications of these experiments are that one needs to be quite careful
about increasing it unless the recruitment policy is also reconsidered. The
real implication is that it is consultants who matter and that ways of increas-
ing the effect of their numbers need to be found .
We referred to 'pet theories' in Fig. 1.3 and someone in the firm argues
that 'we need to get rid of some of the old stagers and bring in some new
blood ', so the reader should experiment by reducing DEMPLOY' Another
theory is that there are better ways of managing projects and of looking
after clients, both of which would have the effect of reducing NNQCPP
from its current value of about 1.8.
The reader should revisit this problem later in the book; for the present
we shall tum to the combat model, which also has some strange behaviour
and illustrates some more of the features of system dynamics modelling.
THE SIMPLE COMBAT MODEL
In Chapter 3 we studied a simple model of two forces in combat. Although

military modelling with system dynamics has been a highly successful appli-
cation area, the reader may not feel particularly interested in the topic.
Nevertheless, this model has many interesting features of general appli-
cability, so the effort of studying it will be worthwhile. The influence
diagram is reproduced in Fig. 6.8.
3 The second part of Fig. 6.7 is produced by model FIG6-7M .COS on the disk, which has been
set up to produce the comparative plots on the same page, purely to make the COSMIC plotter
match the layout of the book.
,000 eoo 500 ,50
1: 700 1141 450 i
1
7001600~ 400i 125
000 : -HI 350i 100

1500' 300.,1
: I
200~ 4oo~ 250~
Ji
75
3OO11!1 2oo~ 50
150~
: 200-iJ '
-200 , J I'I 1oo~ 25
i 100ij 50~
I
-&cx,j
, 0 JJI' oj
, 0
20 . 30. ~o . 50. Bo . 70 . Bo . 90.
I !\ ! L'esc TIME (MONTH) SllIJLATED TIME
I, II,~--==::~e~S'fI (PRO.JECTS/MONTti) BASIC OEMAHO FOR CONSlLT4NCY

(PRD.... CTS) CIRlEHT HUMIl£A OF PRD..lf:CTS
(PEOPLE) HUMIl£A OF QUALIFIED COHSU..TAHTS AVAILAlll.E
L ..~=:.==--:- ~mEF
(PEOPLE! TRAINEE CDHSULTANTS BEING TAAINED
(PRD.... CTS) NET DEFICIT BETWEEN BDC ANO ROHP
A - HIGHER PRODUCTIVITY
'000 eoo 500 ,50
1: 7004111 450i
1
7°°~1' 600 i! 400i 125 /'
0001' 500-H
:
l 300.,"
I
~ ."
100
~/
/-------_/ -------/'_-- -
/'---"'---,
/ / " , - .....
//
J
2ftO~ 400~ 250~ 75 _ .... ---- --- ...., /' -,-, /'
300 '1 2oo~ 50 " '- / / , --,
i: 2oo-iJ 150i'
11
-2&0 J I'1 100i I 25
-eoe.i
11oo~
oJJ oJ
50
1 0
o. 30. ~o. 50. Bo. 70. Bo. 90 . '100. I, 10. ~20 .
! !III i L'eoe TI ME (MONTH) SIMUU TED TIME
(PAD.JECTs/MDNTHI BASIC DEMAND FDA CQNSlA. TANCY
:I c-==.== ~g~~T
------NPAOJ (PRQ..lf:CTS) C~NT NUMIl£A OF PAD..ECTS
(PEOPLE) IIJNIIEA OF QUALIFIED COHSU..TAHTS AVAILABLE
(PEOPLE) TRAINEE COHSU..TAHTS BEING TRAINED
L - .. _ •. _ . •_ .. _ , NETDEF (PRQ..lf:CTSI NET DEFiCIT lIE_EN soc AND AONP
B - LOWER PRODUCTIVITY
Fig.6.7 Counter-intuitive consultants.
The defence planning problem

The purpose of the model is to help Blue identify the 'pressure points'
where investments in assets or changes in the commander's intentions for
fighting the battle might affect Blue's chances of winning and thereby, one
hopes, deter Red from attacking. In practice, Blue's government will have
a limited amount of money to spend on defence, so it will be interesting
to see where money can be spent to have the greatest effect on Blue's
/-------_/ ......... --------_/

/'"
'--=-/-=------------------------------------ -- - ----- ------
I188m TIME (MONTH) SIIU.ATED TIME

A - BASE CASE PAOOUCTIVITY
..
B - HIGHER PRODUCTIVITY
~ ·QCAPT C - LOWER PRODUCTIVITY
QCAPT (PEOPLE) TOTAL QUALIFIED CONSULTANTS (c)

1000
Iii
5oo-Hl
250
I
I
o
I
- 250
I
-500
. 50. Bo. 70 . Bo. boo 'lOb . ~ 10. ~20 .
11~~tR~~
TIME (MONTH) SIIU.ATED TIME
A - BASE CASE PAOOUCTIVITY
8 - HIGHER PRODUCTIVITY
~ ·Nh~EF C - LOWER PAOOUCTIVITY
NETOEF (PROJECTS) NET DEFICIT BETWEEN BDC AND RONP (d)
Fig. 6.7 Continued
defence capabilities. We shall test this below and again when we study
optimization .
The model equations

As usual, the equations are shown in groups, each followed by a brief
explanation. However , only some of the groups are explained , and the
read er is invited to write the rest. The definitions appe ar at the end of the
equations and the reader should note the distinction made between, say,
-~.
Acc,,-- foI<;. -- --- ---- .... ....
-,
•
Force Ratio .-..
,
\ (Red/Blue)
\ " \
\
- \
\
\ '"
/ ",, "
--- t-
I
\
1_
_$hoi 11«11la1.
or,.. \ \
\ /'" \ I / " ..... \
\ /'
\ I : I
\0 +1 \ I ToldRed + I
Force Ra1lo
Trigger \ I / '"Firing Rate "' - ...... I
I \ I / ' , / ,- - H- - ....
I \ I I \ I I '
I \ I I \ I I \
I \
I ~I. ,/ • ~ - \
I B1U:lOS1
,, c~~ .:-- ~=: ....-eR---C~e~=~Rote • RR~=~es
I \ Rote
, f
I I
-c -, /
'r " /
/ I
/ /
""
I / - - ; 1- - .... " \
-,
1 I ' -, '~
I \ ,
I
, / \
Blue Re38N9S _ Blue Reserve ~ Blue peeerve ------+ ' . Blue in Red l oss
I RerT'IOirling . - - - Commitment Rate ~ ArcivatRate Combat Rate
\ +;r t I +~ \
\ / I I I \
\ ", / I \ I \
",'" /'
, .... \ / \
---- / \ I \
\ / \
-c.:='ff~ '..:. Toto/Blue _ / \
firing !lole '-/BA»
l SIloI
'IueRd. ./
«F.".-
....
Fig. 6.8 The simple combat model.
Red men and Blue men. This is done to make dimensional checking as
rigorous as possible. The full model appears on the disk as FIG6-9.COS .
* Figure 6.9 The Simple Combat Model
note
note
note start of war
note ======
note
a wswitch.k=step(1 ,5)
It is often good practice in system dynamics models to prevent anything
happening in the first few time periods. This helps to ensure that the
dynamics subsequently produced are genuinely those of the model's rela-
tionships and not due to faulty initial conditions . In this case, WSWITCH
allows the battle to start when TIME=5.
note
note blue force sector
note =========
note
note blue forces in combat
note
I blue.k= blue.j + dt*(brar.jk- blr.jk)
n blue=iblue
c iblue=10000
Blue forces in combat are a simple level. The initial value is set to a constant
to allow different balances of force distribution between front line and
reserves to be tested easily.
note
note blue losses
note
r blr.klvtrfr.k-bprs
This is what is shown in the influence diagram, Fig. 6.8. Note that Total Red
Firing Rate, TRFR, is modelled as an auxiliary, not the rate which one
would expect. This is another case of the flexibility of the software allowing
one to have two options . Since there is, in the simple model, no stock of Red
ammunition to be depleted, there is no need for TRFR to be a flow rate.
When the stock is modelled, TRFR COUld, and should, be a rate, and the
equation would be
r blr.kl =trfLkl*bprs
which again breaks the strict rules of the connections between levels, rates
and auxiliaries which were shown in Fig. 4.1. It is important to grasp that,
though the rules can be broken, one should only do so after careful thought!
note
note blue reinforcement arri vals
note
r brar.kl=delay3(brcr.jk, bmdel)
c bmdel=6
note
A simple flow from Reserve Commitments.
note blue reinforcement commitments
note
r brcr.kl =min(blres.k/dt,bmcap)*frtrig.k
a frtrig.k =clip(l,O,frat.k,bcafr)*wswitch.k
c bcafr=2.0
c bmcap=4000
Blue Reserve Commitment Rate depends on the Blue commander wanting
to commit reserves, which only happens when Force Ratio, FRAT, exceeds
a desired level, thereby making the Reserve Commitment Trigger change
to 1. Note the use of the non-negativity constraint, BLRES.K/DT. Note also
the use of auxiliaries to break the equations into simple components which
are easier to check than one complicated equation for BRCR. The use of
WSWITCH prevents Blue from initiating the provocative act of commit-
ting reserves before the war has started, even if his front line troops are
outnumbered.
note
note blue reserves
note
I blres.k=blres.j+dt*(-brcr.jk)
n blres=iblres
c iblres=20000
A simple drain of Blue reserves as they are committed.
note
note blue firing rate
note
a tbfr.k=blue.k*brfpm *wswitch.k
c brfpm=150
c rpbs=O.OOl
The total rate at which the Blue force fires depends on how many troops
there are and how fast each man fires, once the war has started.
note
note force ratio
note ======
note
a frat.k =red.k/(blue.k + beps)
c beps=0.001
The equation for Force Ratio illustrates the need for protection against
division by 0 if BLUE.K becomes O. The use of a small number, such as
BEPS, is a useful trick.
note
note red force sector
note ========
note
The reader should now write the equations for the Red force, recalling that
Red 's 'policy' is simply to commit reserves as fast he can, once the war has
started. Use the following numerical values.
c ired = 15000
c rmdel=6
c rmcap=2500
c irres=500oo
c rrfpm=150
c bprs= .00033
note
note output and control
note ==========
note
c pltper=0.25
c prtper=lO
c length=60
c dt=0.25
print 1)blue,blres
print 2)brcr,brar
print 3)blr,rlr
print 4)red,rres
print 5)rrcr,rrar
print 6)frtrig
plot blue=a,red=b(0,60000)
plot blres= a,rres = b(0,6oo00)/frtrig=c(0,2)
Although the format is used in some of the system dynamics packages, but
not all, the principle of printing all the variables so that the model can be
checked is universal. The plotting interval, PLTPER, is set to be the same
as DT to give the smoothest possible curves in Fig. 6.9. The scale for
FRTRIG is set to force it to be in the middle of the page, where it will be
easier to see.
note
note ============
note
run Base Case
A simple test to see how well the model works.
note
note definitions of variables
note = = = = = = = = = = =
note
d bcafr=(red/blue) blue commander's acceptable force ratio - the
x force ratio at which he commits reserves
d beps= (blue) small number of blue to prevent division by zero
d blr=(blue/hour) blue loss rate
d blres=(blue) blue reserves remaining
d blue=(blue) blue forces in contact with red
d bmcap=(blue/hour) blue movement capacity
d bmdel=(hour) delay in before blue reserves committed to combat
x become effective
d bprs=(blue/rshot) blue killed per red shot - the red standard of
x marksmanship
d brar=(blue/hour) rate at which blue reserves enter battle
d brcr=(blue/hour) rate at which blue reserves are committed to battle
d brfpm=(bshot/hour/blue) blue rate of fire per man
d dt=(hour) simulation time interval
d frat=(red/blue) force ratio of red: blue
d frtrig=(l) switch to trigger commitment of blue reserves
d iblres= (blue) initial blue reserves
d iblue=(blue) initial blue force in contact
d ired=(red) initial red force in contact
d irres = (red) initial red reserves
d length=(hour) simulated period
d pltper=(hour) plotting interval
d prtper=(hour) printing interval
d red=(red) red forces in contact with blue
d rlr=(red/hour) red loss rate
d rmcap=(red/hour) red movement capacity
d rmdel=(hour) delay before red reserves committed to combat
x become effective
d rpbs=(red/bshot) red killed per blue shot - the blue standard of
x marksmanship
d rrar=(red/hour) red reserve arrival rate
d rrcr=(red/hour) rate at which red reserves are committed to combat
d rrcsw=(l) switch to control commitment of red reserve s

d rres=(red) red reserves remaining
d rrfpm=(rshot/hour/red) red rate of fire per man
d tbfr=(bshotlhour) total blue firing rate
d time=(hour) simulated time
d trfr =(rshot/hour) total red firing rate
d wswitch = (1) switch to denote onset of combat
Testing the model

The behaviour in the base case is shown in Fig. 6.9, in which the scale values
for RED, BLUE, BLRES and RRES are to be multiplied by 1000 to get the
actual values. The plots show that something is very seriously wrong with
this model'. In the upper graph , the error is that BLUE reaches about
-100000 and RED goes to about 160000, despite having started with only
65000 in total.
The behaviour in the lower graph is a little more plausible. The war starts
at TIME=5, at which point Red starts to commit reserves as fast as he can.
Since he has 50000 reserves and a transport capacity of 2500 men/hour, it
should take 20 hours to commit all reserves , and it does. Similarly, it is also
correct that Blue commits all his reserves in 5 hours. Both sets of reserves
stop at 0, so the non-negativity constraint is working. The dynamics of
FRTRIG are strange. Blue starts to commit reserves at TIME=12 and
continues to want to do so after TIME=16, even though there are no more
to commit. At about TIME=34, Blue suddenl y stops wanting to commit
reserves, which is odd.
The reader must pause and try to work out what is wrong before going
any further. To do so, run the model , look at the output, and think.
The error is that we allow Red and Blue to continue inflicting casualties,
even when all the opponents have been wiped out. In other words, there is
no non-negativity constraint on Blue Loss Rate , or on Red 's, since there is
a possibility that Blue might be the victor.
The corrected model is FIG6-1O.COS, the behaviour of which is shown in
Fig. 6.10. Alter the model , and the diagrams in Figs. 6.8, 3.8 and 3.14, to get
this behaviour.
Figure 6.1O's dynamics are much more plausible. Blue starts to suffer
casualties when the war starts and, by TIME=12 is outnumbered by 2:1.
At that point, the Blue commander starts to commit reserves and the curve
for BLUE starts to rise a few hours later. Unfortunately, there are insuffi-
cient reserves to turn the tide, and Blue is wiped out by about TIME=33.
Red 's remaining reserves continue to arrive on the field, and RED ends
4 Congratulations to the reader who spotted the deliberate error in the model and didn't get
the behaviour shown in Fig. 6.9!
200-n
I
150~ I
I
I
l00~
I
/
/
/
/
/
./
------ ---- ----"

50
o ------------- -------
-50
TIME lHOUR) : I4l.\.ATEQ TINE

(Ill.UE) BLUE FORCES IN CONTACT NITH REO
(a) (RED) RED FORCES IN CONTACT NITH Ill.UE
2'1
1.7!1J 50. ---,,
1.~ ,,
,,
us] 40 . ,,
,,
I ,
1.
1 30. ~------]
. 7~ 20 ..-t+-
I '\\
-{,
II
.~
I \ I
.24 10 .
I " I
o.J
I \
0 .:-+,-~t._....._lmi____lon::'__4m____'4,....._.........
I
I 1110~ ' . ..L_l".,.__l..rror___l;.",.___1....___6.___{,
o.
L~R~R~S
TIME lHOUR) SII.....ATEo TIllE
I~---FRT~IG IIll.UEI BLUE RESERVES REMAINING

lREDI RED RESERVES RENAINING
(11 SNITCH TO TRIGGER ClMlITIlEN T OF BLUE RES£AVES
(b) BASE CASE
Fig. 6.9 Output from the simple combat model.
the battle with a total strength of about 33000 men. The lower graph shows
the previous behaviour for RRES and BLRES , but FRTRIG, having
switched on at TIME=12, stays on, showing the Blue commander's
wish to commit reserves continues even after the battle is lost, which is
plausible.
Unfortunately, even though the graphs look right, the model still contains
errors. Run the model, looking at the tables and, perhaps, plotting a few
more variables to see if you can detect the error. Plotting all the variables is
usually just confusing: the tables of output are more informative.
---------------------
/
/
/
/
.>
---,,
__ .......
/
/
.,.,..-.........
/
....
TI ME (HOl.fl) SINULATED TIME

(BLUE) BLUE FORCES IN CONTACT WITH RED
(RED) RED FORCES IN CONTACT WITH BLUE (a)
2'1 60 .
1.7~
50 . ---,
1 .~ \
\
\
\
1.2sJ 40. \
I.,
\
\
I 30 .
\
\
\
~---------------
1\,
. 7~ 20.+t------{
\
\
\
\
.5.j \
\
\
.2~
\
10. \
\
\
\
\
O.J O.::'-J.r--l,...,..... \
I L' .
* 10 i TIME (HOUR) SI .....ATED TIME
L_~~R~S
I----FRy~IG (BLUE) BLUE RESERVES R....INING
(RED) RED RESERVES REMlINING
(n SNUa. TO TRIGGER COMMITMENT OF BLUE RESERVES
BASE CASE (b)
Fig. 6.10 The corrected combat model.
The error is that Red still continues to fire even after Blue has been
completely wiped out. In other words, TRFR does not go to zero when
BLUE does. In a sense , this does not matter, as neither Red 's nor Blue' s
ammunition stocks feature in the model. It is, however , a logical error and
one can have little confidence in a model which contains errors, let alone
looking rather foolish and destroying the client 's confidence if he or she
happens to question that aspect of the model. FIG6-lOM.COS on the disk
contains the final corrected model.
We shall stud y this model in more det ail in Chapters 7 and 9, but a few
final comments are called for.
In the first place, this model differs from the two previous examples in
that it has no external driving force. The first two models are driven , but this
one is closed and its dynamics arise entirely from its initial conditions and
structure.
Secondly, although non-negativity constraints need to be added to the
Blue and Red Loss Rates in Fig. 6.8 they are not needed for Blue Ammu-
nition Reserve Remaining, Blue Ammunition Stocks, Engineer Supplies
Remaining and Long Range Ammunition Stocks in Fig. 3.145 • For the first
of those, the problem statement implied that they were so large that they
could never be used up. In fact, they are probably irrelevant to the problem
and could be omitted. If they are retained, it would be as well to add a
constraint, partly for consistency but also to be on the safe side. The other
three all work by affecting the rate at which they are used; for instance Blue
Ammunition Stocks influence the Blue Rate of Fire per Man and, when
stocks reach zero, consumption automatically ceases. It would not be wrong
to add the constraint, but it would be an unnecessary complication, which is
bad practice. A model should be kept as simple as possible.
THE PREDATOR/PREY MODEL
Predator/prey systems, such as that shown in Fig. 3.5, and reproduced in

Fig. 6.11, present some interesting modelling problems. They are usually
'closed' or undriven models in which the dynamics arise entirely from the
structure, initial conditions, and parameter values (usually including sharp
non-linearities) and there is no external driving force, as was the case for
the washing machine company and the management consultants. They also
differ from the closed combat model examined earlier in that the dynamics
cannot be prevented from taking place in the early stages of the simulation,
as the war switch allowed us to do. That means that it is difficult to check for
false dynamics due to faulty initial conditions or errors in the equations. To
add to all that there are usually no 'managers' to tell us what the system is
supposed to achieve.
The purpose of the model

The purpose of the model is to represent the dynamics of predator/prey
behaviour and to study how those dynamics change with uncertainties in
the data. When the system has been well understood, the purpose might
develop into understanding how the system might be managed by human
intervention. For instance , foxes might have to be culled and wild rabbits
are a delicious food.
5 The point is not only to do with military modelling; it could arise in any other type of
problem .
NotmQ/ ,. -- - - - - - , -,
Fox /
-,
F_I/y
FOXFEe "- ,
I
I
-\
I
I
I -:
I /
I /
~ .;+
Fox
Conceptlon
Role
feR
\~
~~ Rate
FBR
Fig. 6.11 The predator/prey system.
Developing the equations

You will probably be able to write most of the equations for yourself if you
have assiduously studied the previous examples. It is suggested that you try
to do so, concentrating on the equations and leaving the numerical values
blank for the moment.
The following suggested solution contains roughly plausible numerical
values which produce some interesting dynamics. After studying the solu-
tion and its behaviour the interested reader may care to try to find some
more 'accurate' data to see what effect they have.
* Figure 6.12 The Predator/Prey Model

note
note
note rabbit population sector
oo~ ======= = =====
note
note rabbit popul ation
note
I rabpop.k =rabpop.j + dt*(rabbr.jk - rabkr.jk)
n rabpop=inrab
c inrab =300
note
note rabbit reproduction
note
r rabcr.kl =rabfec*rabpop.k
c rabfec=O .1
r rabbr.kl =delay3(rabcr.kl,drgest)
c drgest=4
note
note fox population sector
note == = === == = = =
note
note fox population
note
I foxpop.k=foxpop.j +dt*(fbr.jk -fdr.jk)
n foxpop=ifox
c ifox=20
note
note fox reproduction
note
r fcr.kl= foxfec*foxpop.k
c foxfec=O.05
r fbr.kl =delay3(fcr.jk,dfgest)
c dfgest=7
The equations so far are probably self-evident, though the reader will wish
to check them .
note
note fox predation of rabbits
note
r rabkr.kl =fheff*rabpop.k*foxpop.k
c fheff=O.005
The Rabbit Kill Rate is modelled by equations which are also used to model
the spread of an infectious disease from the infected to the uninfected parts
of the population. The parameter for Fox Hunting Efficiency, FHEFF, has
the dimensions [[lIweek]/fox] which is the fraction of the Rabbit Popula-
tion which a fox can kill in a week.
The underlying idea is that R rabbits occupy an area of X square miles. A
fox can search Y square miles per week so the number of rabbits it encoun-
ters per week will be R XY/X. If the conditional probability of a rabbit
being killed, given that it has been encountered, is P, then the weekly rate
of rabbit kills per fox will be PX(Y/X)XR The Fox Hunting Efficiency,
FHEFF, is PX(Y/X). In other words, FHEFF is the fraction of the rabbit
population which a fox will encounter in a week of hunting, multiplied by
the fraction of those encounters which end in a fox killing a rabbit.
note
note fox food supply
note
I akr.k =akr.j + (dt/fatime )*(rabkr.jk -akr.j)
n akr=rabkr
c fatime=2
a avipf.k=akr.k/foxpop.k
The Rabbit Kill Rate is smoothed over a time constant which represents the
period for which a fox can go hungry before hunger starts to have an effect
on its lifespan.
a rlife.k=rabpop.k/(akr.k +reps)
c reps=O.OOOOl
It will be useful to have an indicator of the average lifespan of a rabbit. One
could devise a second smoothing time to calculate that, but we use
FATIME for simplicity. REPS is introduced to prevent division by zero if,
for any reason, AKR reaches zero.
Although most of the software packages would accept it, it would be
logically incorrect to write the equation as
a rlife.k=rabpop.k/(rabkr.KL+reps)
RABKRKL is the kill rate which is about to occur in the impending DT , so
it is wrong to calculate RLIFE.K, something which exists now, by using
something which has not yet happened, which is what the .KL postscript
represents.
RABKRJK cannot be used, because the time-shifted window has moved
to TIME.K, so the .JK values no longer exist. In any case, RABKRKL may
vary pretty sharply and it would be better to get a smoother view of RLIFE.
note
note food supply effect on fox lifetime
note
a fratio.k=avipf.k/nfrpf
c nfrpf=5
a fsrf.k = tabhl(tfsrf,fratio.k,O,1.0,0.2)
t tfsrf=0.O/O.024/0.12/0.8/0.98/1
This is the formulation discussed in Chapter 5.
note
note fox death rate
note
r fdr.kl =foxpop.k/(dlife.k +minlt)
c minlt=1
a dlife.k= nfltf*fsrf.k
c nfltf=120
Finally, Fox Death Rate is modelled using a first-order delay with fox
lifetime dependent on the food supply. It would have been easy, though
somewhat more laborious, to have written explicitly the equations for a
third-order delay (recall Fig. 4.8), which would have allowed us to distin-
guish between immature, mature and old foxes, probably with different
tables to reflect the possibility of old and immature foxes dying of hunger
more easily than mature ones . Whether that would add much to the cred -
ibility of the model is a matter of judgement. It would certainly impose
greater demands for numerical data.
note
note control and output sector
note ========== ====
note
c dt=0.125
c prtper=5
c pltper=1
c length = 300
print 1)rabpop,rabbr
print 2)foxpop,fcr,fbr,fdr,
print 3)rabkr,akr,rlife
print 4)fratio,fsrf,dlife
plot foxpop = a(0,40)/rabpop = b(0,800)/fdr=c(0,4)
plot dlife=a(0,120)/rlife= b(0,40)
Choosing a value for DT requires us to evaluate two candidates. The third-
order delay for Rabbit Gestation is 4 weeks long, which would require
DT <4/12. Fox Lifetime is a first-order delay with a minimum value of 1 if
there is no food at all for foxes. That would require DT <114. To be on the
safe side, we put DT=0.125, half the value required by MINLT.
note
no~ ============ =
note
run A - Trial of Model
c inrab=600
run B - Larger Initial Number of Rabbits
Two experiments, to test the effects of uncertainty about the numbers of
rabbits in the area.
note
note variable definitions
note ===== = = = =
note
d akr=(rabbits/week) average kill rate of rabbits
d avipf=«rabbits/week)/fox) average food intake per fox
d dfgest=(week) average length offox pregnancy
d drgest=(week) average length of rabbit pregnancy
d dlife=(week) average fox lifetime
d dt=(week) solution interval
d fatime=(week) averaging time for fox food intake
d fbr=(fox/week) fox birth rate
d fcr=(fox/week) fox conception rate
d fdr=(fox/week) fox death rate
d fheff=«l/week)/fox) fox hunting efficiency - the proportion of the
x rabbit population that a fox will kill during
x a week
d foxfec=(l/week) fox conceived per fox per week
d foxpop=(fox) fox population
d fratio=(l) food intake of faxes relative to normal
d fsrf=(l) reduction factor on fox life span due to relative availability
x of food
d ifox=(fox) initial number of faxes
d inrab=(rabbits) initial rabbit population
d length=(week) simulated duration
d minlt = (week) minimum lifespan of faxes
d nfltf=(week) normal lifespan of a fox when food supply is normal
d nfrpf'<r(rabbits/weekj/tox) food intake per fox for normal life span
d rabbr=(rabbits/week) rabbit birth rate
d rabcr=(rabbits/week) rabbit conception rate
d rabfec=(l/week) rabbits conceived per rabbit per week
d rabkr=(rabbits/week) rabbit kill rate
d rabpop=(rabbits) rabbit population
d reps=(rabbits/week) small number to prevent division by zero
d rlife=(week) average lifetime for a rabbit
d tfsrf=(l) table for food supply reduction factor
d time=(week) time within simulation
BOO, 40.
I
I
700.J 35 .
I
I
I
600-J 30 .
I
I
500"': 25 .
I
I
400~ 20 .
I
I
300~ 15.
I
I
200 .J 10 .
I
I
I
100-J 5.
I
I
oj O.
60. ~50 . 11BO.
TIME (MEEK) TINE MITHIN SIMULATION
(FOX) FOX POPUlA nON

(RABBITS) RABBIT POPULATION
(FOX/MEEK) FOX DEATH RATE
40 ., 150
I
I
35 .J
I 125
I
30 .J
I
I 100
25 .~
1
1
20 .~ 75
I
I
15 .oj
I 50
I
10 .J
1
I
25
5.JI
I
O.J 0-+r---1",.,-i;o-:- bOo 1120 . 1150 . ~BO . 210 . 240 . 210 . ~OO.
TIME (MEEK) TIME MITHIN SIMULATION
(MEEK) AVERAGE FOX LIFETIME
(MEEK) AVERAGE LIFETIME FOR A RABBIT
A - TRIAL OF MODEL
Fig. 6.12 The predator/prey model.
Testing the model

The model is on the disk as FIG6-12.COS, and the behaviour for two
different initial values of Rabbit Population is shown in Fig. 6.12. The
behaviour is characteristic of predator/prey models in that the two
BOO, 40.
I
I
700...1 35 .
I
I
600J 30 .
I
I
500~ 25 .
I
I
400~ 20 .
I
I
300~ 15.
I
I
200...1 10 .
I
I
100J 5.
I
I
oj
TIME (MEEK) TIME MITHIN SIMULATION
(FOX) FOX POPILUION
(RABBIT9) RABBIT POP\A.A TION
(FOX/MEEK) FOX OfATN RATE
40 ., 150
I
I
35 .J
I 125
I
30 .J
I
I 100
25 .~
I
I
20.~ 75
I
I
15.~
I 50
I
10 .J
I
I
25
5 .J
I
I
O.J 0 *--l."....--\:,-,.,..--i,,,,-~,......,...--I6r-b..----~,,..----6~*'m---l,00 .
i J·~l:!~~
TIME (MEEK) TIME MITHIN SIIOA.ATION
(MEEK) AVERAGE FOX LIFETIME

(MEEK) AVERAGE LIFETIME FOR A RABBIT
B - LARGER INITIAL NUMBER OF RABBITS
Fig. 6.12 Continued
populations show sustained oscillations. It must be emphasized that the

oscillations arise entirely from within the model and are not due to outside
influences.
In the first case, the upper half of the graph suggests that the 300 rabbits
initially present are not sufficient to support a population of 20 foxes, so the

Fox Death Rate is relatively high. The fall in Fox Population allows rabbits
to live long enough to breed prolifically and Rabbit Population rises. Natu-
rally, the increased availability of food causes a rise in Fox Population, and
so on. The period of the oscillation is about 70 weeks and the slight irregu-
larities in the curve for Fox Death Rate are due to the changes in slope in
the piece-wise linear approximation in the TABLE function .
The lower half shows that the average lifespan of both species fluctuates
considerably, but at the times which one would expect from the rise and fall
of the respective populations. The lifespans at the troughs are rather short,
probably less than the time required to become capable of breeding. At
these points , rates of conception are low, so the error is not large. It would
be easy to complicate the model to prevent breeding from happening when
the average lifespan is less than a certain value.
With 600 rabbits, the pattern is similar, but in reverse. The model starts
with ample fox food , so Fox Population rises, after which the same pattern
of oscillation sets in, with the same period. The contrast between Figs.
6.12A and 6.12B suggests that there might be some initial number of rabbits
which would stabilize the system. We shall examine that topic in a moment,
but first we must test the model in another way.
The value of DT was set to 0.125, but, if we run the model with 300 rabbits
and successively set DT to 0.25, 0.5 and 1, we obtain the comparative results
shown in Fig. 6.13 (FIG6-13 .COS on the disk) . The behaviour patterns with
DT=O .125 and DT=O.25 are not appreciably different, though there is a
slight indication that the oscillations become larger with DT=0.25. With the
two larger values of DT, the oscillations certainly become greater, to the
point where RABPOP goes off the graph scale. The period of the oscilla-
tion also lengthens.
Clearly, the value of DT has an effect on the results, and this arises from
the assumption, discussed in Chapter 4, that DT is so small that the rates
can be assumed to be constant during DT. This is called Eulerian integra-
tion . Some of the system dynamics packages use other integration methods,
but, in all cases and for all models, it is essential to double and halve the
value of DT initially selected" to see if it makes any noticeable difference to
the results from the model. The reason for this is that DT is chosen, by the
modeller, to get the equations which purport to represent the system to run
on a computer. It has nothing to do with the real system, so its value should
not significantly affect the behaviour of the real system as simulated by the
model. If DT has to be reduced to an absurdly low value in order to get
invariant behaviour there is something seriously wrong with the model.
6 Recall that the acceptable values of DT are 0.5, 0.25, 0.125 etc. There are the decimal
equivalents of exact binary fractions and are used in order to avoid rounding error when
functions such as PULSE and STEP are activated by comparison with the imbuilt TIME
variable.
40 ,
35,
10 ,
5, iI
I
o , .1.-"""*-~-i.1'r-;."....--~.,.,--lno"...-""=r.,....-~--";',,.,.,.-----"",.,.,-----l,
, 00.
TIME (OffK) TIME O!THIN SI .......TION
!II!--F
L'F8 XP8P
I
A - DT-O •• 215
XP P 8 - OT-O,28
C - OT-0.8
L.:=~8~~8~ o - OT-1
FOXPOP (FO X) FOX POPULATION (a)
BOO
700
600
400
II L' , TIME ' (OEEK) 'TIME OI~IN SIl.~'TION

, 00,
' 1 RABP~P •8 -- OTo(), '28

1 -- RASP P OT-O,28
L""::: ~~G~ ~ g: g~:~,8
RABPOP (RABBITS) RABBIT POPULATION (b)
Fig. 6.13 DT in the pred ator/prey model.
The pattern of oscillation in Fig. 6.12 implies that there might be popula-
tion levels which make the system stable. A little experimentation shows
that, with 411 rabbits and 20 foxes initially, the populations do stabilize,
though with rather short lifespans. This is shown in Fig. 6.14, the model
being FIG6.14.COS on the disk. With 300 rabbits, there is no value of fox
population which stabilizes the system. One might conclude that fox culling
will not stabilize the system and that the correct policy is to introduce
enough rabbits to bring the population up to 411 at a time when there are
40.
35 .
30 .
25.
20 . ------------------------------------------ ---
15.
10 .
TIME (WEEK) TIwE WITHIN SIwu..ATION

(FOX) FOX POPULATION
(RABBITS) RABBIT POP\A.ATION
(FOX/WEEKI FOX DEATH RATE
40 ., 150
I
I
35 '1
I
125
30 .J
I
I 100
25)
I
I
20 .~ 75
I
I
15 ·i
I 50
I
10 .J
I
I
25
5.JI
I
O.J o-h----I,-"r-""r---hrr-,..,rn---urn---u,.,.,---ln,.,.,---ln,.,.,---l,-"..----I,OO .
L' TIME
i--------. RLIFE
I (WEEK) TIME WITHIN SIMll.ATION
DLIFE (MEEK) AVERAGE FOX LIFETIME
(WEEKI AVERAGE LIFETIME FOR A RABBIT
(b) 411 INITIAL RABBITS
Fig. 6.14 Stable populations in the predator/prey model.
20 foxes present. This is an example of the way in which system dynamics

models generate novel insights into problems.
The dynamics of a non-linear system are heavily dependent on its initial
conditions . We saw that the behaviour when INRAB=300 is rather differ-
ent from that with INRAB=600. When INRAB is set to 30000, the dif-
ference in behaviour becomes very dramatic, as seen in Fig. 6.15
(FIG6-15.COS will produce that illustration). The cycles are still present,
but the period has extended to about 150 weeks and the peaks of the cycles
5'1 Boo
l
40 .
4.5.j 700~ 35 .
4. I :
1 600-1 30.
3.J :
I 5OO~ 25.
3'1 :
2. 5.j 400~ 20 .
I
d 3oo~ 15 .
1.,1 200-l: 10 .
1'1 I
I
.s.j 100~ 5.
O. J 0 J 0. *-~.....-~...-=I;;;;;::.J...,.,.,--~rl~=l;,;"""'..,.".--l",,.,r.lJ;;;,:;=:r,,.,.,.,--l;.,,j.,
~·~~~~g~
TIME IWEEK) TIME WITHIN SIMU..ATION
I i (FOX) FOX POPUU TION
(RABBITS) RABBIT POPULATION
------FOR (FOX/WEEK) FOX DEATH RATE (a)
40 ., 150
I ,,
I
,
I
I
1
I
I , ,,
I
35 .-j
I
125
I
I I
I ,
30 .J
I
I
I
I
I
I
I
,
I
I
100 \
I I
25 .~ I
I I I
I I I I
,
I I \ \
I \ I
20.~ 75 I I
I \ \ \
I \ I \
\ \ \
15 .~ I \ \
I 50 \ \ \
I \ \ \
10 ..J \ \ \
I \ \ \
I \ \ \ \
25 \ \
5.JI \ \
\ \
,,
\ \
,
\ \
\
"
I
\
\ , \
\
O.J
I
"
I L" TIME (WEEK) TIME WITHIN SI .......TION
'--------- ~l:i~~ 1::1 :~~: ~~:ri~r:T~: AFWlBIT

20 FOXES - 30000 RABBITS (b)
Fig. 6.15 Larger initial numbers in the predator/pre y model.
ste adily decay. If the model is run for 5000 days the cycles do reach approxi-
mate stability.
The behaviour is so strange that a littl e further inve stigation is called for.
It turns out that RABKR reaches a maximum valu e of 3616 rabbits per
week , which means that th e average fox is killing 180 rabbits per week, or
26 times the food intake it requires for normal life. It is possible that th e
model should be modifi ed to prevent a fox killing more than, say, 10 rabbits
per week. What would th e effect of that be with this larg e initial population
of rabbits?
The techniques used here , particularly the modelling of Rabbit Kill Rate as
dependent on the interaction of both populations, are quite widely appli-
cable. A model of this general type might, for instance, be used to study the
practicalities of reindeer harvesting when there is also a population of
wolves to manage. The same general approach can be used in modelling the
spread of a disease. Similar ideas are used in some types of combat model-
ling, such as air defence and submarine warfare.
THE RULES FOR TIME POSTSCRIPTS FOR VARIABLES
In Chapter 4, the discussion of the fundamental equations of system dynam-

ics stated some rules for the postscripts which are permissible on the right-
hand side of an equation for a given variable type on the left-hand side. In
essence, these are:
L ALEV.K=ALEV.J+DT* (RATEs.JK)
L SLEV .K=SLEV.J+ (DTrrCON) (RA TE.JK-SLEV.J)
R RATE.KL=f (LEVs .K,AUXs .K,PARAMETERS)
R RATE1.KL=DELAYn (RATE2.JK,DEL)
A AUX1.K=f (LEVs .K,AUX2s.K,PARAMETERS)
There are two possibilities for a level: ALEV is a pure accumulation of a
physical flow, while SLEV is the smoothing of information about a physical
flow. Similarly, there are two possibilities for a RATE, depending on
whether or not there is a delay in the equation.
The lower case 's' shows that there may be more than one of the indicated
variable , and numbers, such as AUX1 and AUX2s , suggest that AUX1 can
depend on a number of other auxiliaries, though not, of course , on itself.
The postscripts in bold are those which Fig. 4.1 and the fundamental equa-
tions of system dynamics permit on the right-hand side. Strictly, neither
levels nor auxiliaries can appear on the right-hand side of a level equation.
In practice, however, and depending on the software package, these strict
rules can usually be broken where to do so makes it easier to write equa-
tions. We have seen some examples of this in the models studied in this
chapter. Figure 6.16 shows the postscripts which will be accepted by the
software, those in bold being the standard cases, the others involving bend-
ing the strict rules.
The rules for a level must be carefully thought about. If the level is a pure
accumulation, then it is very inadvisable to use anything other than rate
variables within the parentheses on the right-hand side. The level should be
an accumulation of some physical flow, and rates are the variables of
The rules for time postscripts for variables 193
physical flow. The software might well accept a level or an auxiliary, but to
write anything other than a rate may mean that the modeller is seriously
confused about how the physics of the system works and is incorrectly
creating a memory of the physical consequences in it. That becomes even
clearer when we recall the rules for the dimensions of levels and rates given
on p. 88.
If, on the other hand, the level is the smoothing type which creates a
memory of information it is perfectly permissible to have a level or an
auxiliary as the variable being smoothed.
In the manufacturing problem, we used Average Order Rate to represent
the smoothing of the physical inflow of orders. It would, however, have
been perfectly logical to model the smoothed (averaged) backlog or dis-
crepancy using equations such as:
L SBLOG.K=SBLOG.J+(DTrrCON) (BLOG.J-SBLOG.J)
and
L SEPSl.K=SEPSl.J + (DT/TCON) (EPSl.J -SEPSl.J)
Equations such as these could be written, for example, as
L SEPSl.K=SEPSl.J + DT*RCSEPSl.JK
R RCSEPSl.KL= (EPSl.K -SEPSl.K)rrCON
with RCSEPSI representing the rate of change of SEPSl. This would make
a level, SEPSl, depend on a rate, RCSEPSl , which, in turn, depends on a
level, SEPSl , and an auxiliary, EPSl. That is formally consistent with the
Variable on the Right is

Variable on the
Lett is Level Rate Auxiliary
Level J JK: J
(accumulation) I I
-------~------~- -------
I I
Level
(smoothing) J : JK : J
I I
I I
Rate
-------r------r- -------
(no delay)
K I KL I K
I I
Rate
-------,------,--------
I I
(delay) J : JK : J
I I
------~-------r- - ------
Auxiliary K I KL I K
For explanation see text .
Fig. 6.16 The rules for time postscripts.

fundamental equations, but is clearly an unnecessary fuss. The strict rule is

too restrictive for practical use.
For flow rates, we used a rate-dependent rate to model the transition of
trainees to the state of qualified consultant by writing:
R TRQS.KL=TCRKL*PSAT
Again , a rate-dependent rate is not, strictly, allowed by the fundamental
equations. In practice, TCR is a rate dependent on levels and it would be
carrying adherence to strict rules to silly extremes to waste effort by writing
the equation for TCR on the right-hand side of the equation for TRQS and
then multiplying it by PSAT. The equation above is perfectly clear and only
cosmetically violates the strict rules. The main thing in writing equations in
which a rate depends on a rate is to be sure that one is genuinely taking a
legitimate short cut and not using the flexibility of the software to fudge a
real difficulty.
Where a delay is involved, it would, however , be very undesirable to have
LEV.J or AUX.J within the parentheses on the right-hand side. Again , the
reason is that a delay is a physical process, while levels represent accumu-
lations and auxiliaries are usually parts of the fine structure of decision
processe s within the system . It is hard , therefore, to see these as logically
feeding directl y into a physical flow. Again , correct application of dimen-
sional analysis will usually be a great help in avoiding such errors.
In short, most of the system dynamics packages permit flexibility in
writing equations, but the privilege should not be abused without careful
thought.
SUMMARY
This chapter has shown, in careful steps, the building of four models. In
each case, they were tested in various ways, and we shall use some of them
for policy analysis and optimization in later chapters. You should study
these examples carefully and further extend that knowledge by building for
yourself some of the problem models given in Appendix B.
CHAPTER 7
Policy experiments with

system dynamics models
INTRODUCTION
In Chapter 1, we discussed the idea that the main aim of system dynamics
is to develop policies which improve the dynamic behaviour of a system.
This is obviously the intention when the clients are the management of a
business firm, but it is also the objective if the 'client' is the academic
audience for a book , as was the case for the decline of the Maya, studied in
Chapter 3. At first sight, it might seem that the purpose of such a model
would be simply to understand why the system had behaved in a certain
way, but, inevitably, the researcher becomes interested in how the collapse
might have been avoided had the Maya rulers acted differently, and that is,
by definition , policy analysis.
In Chapter 6, we developed models of four problems and showed that,
when the models were tested, their modes of behav iour could be changed
quite noticeably. The behav iour was not necessarily better than the 'base
case', but it was different, and knowing that behaviour can be changed lends
encouragement to the idea that we might be able to design the best possible
robust beh aviour into the system in question.
This chapter will apply that idea to some of the models from Chapter 6.
The aim is not to imply that the solutions to be developed are in any sense
'ideal' or representative of good practice . The intention is simply to show
how models can be experimented with to test policy effects and to indicate
the wide range of behaviour which is usually found. At the end of the
chapter, you should have developed confidence in working with models and
have had your imagination stimulated with ideas about what can be done
with them. In Chapter 9, we shall take this idea of design much further when
we study the use of optimization for policy design in system dynamics
models.
POLICY IN SYSTEM DYNAMICS MODELS
Before we can say much about policy analysis, we should be clear about
what 'policy' means in system dynamics modelling and how policies are
196 Policy experiments with system dynamics models
represented. Since we shall use the Domestic Manufacturing Company,

DMC, for our first policy studies later in this chapter, let us also use it as the
basis for discussing this point.
The Raw Materials Manager orders raw materials on the basis of:
r rmor.kl=eps2.k/tarms+apr.k
The equation says that the Raw Materials Manager has a policy of ordering
raw materials so as, firstly, to keep actual raw material stock up to a target
level, correcting any discrepancies over a period of TARMS weeks, and,
secondly, of replacing the consumption of raw materials in the production
process.
There is no law which states that this must be his policy: it is simply
that which he chooses to use, probably having inherited it from his
predecessor.
An alternative policy would be to order raw materials so as to ensure that
Backlog could be brought down to its desired level within a reasonable time
and to match the current level of Orders. That policy would require the
different equation:
r rmor.kl = epsl.kltarms+ aor.k
These two equations would require two different influence diagrams. Links
would, as it were, have to be switched on and off to depict the two possibil-
ities. When we study policy in the DMC problem, we shall introduce a
notation for doing so.
There are, however , very many other possibilities, and there is no way of
knowing in advance which will give the best performance for the business as
a whole. The only way forward is to experiment with different policies with
the intention of designing that which is most satisfactory.
It is, however , critical to understand that policy has two components. The
first is its structure; the form of the equation and the corresponding set of
links on the influence diagram. The second is the parameters within a given
structure; the numerical value of TARMS might have to be changed con-
siderably in order to get the best results from each of the two options
suggested above.
EXPERIMENTATION VERSUS DESIGN
In the last section we referred to 'experimenting' with models and to 'policy

design', and it is necessary to understand the important distinction between
the two. We shall use 'analysis' to encompass both 'experimentation' and
'design'.
System dynamics models often have many parameters. As we shall see,
the DMC model has seven in its basic form, and it will have many more by
the time we have finished with it. When we introduce structural variations
into the model, the number of parameters increases yet again, and the
number of combinations of those parameters is very large, even though
this is a fairly small model. A large model will have so many parameters
and structural possibilities that the number of combinations becomes
enormous.
With such large numbers of options, coherent analysis becomes very
difficult, so policy analysis usually proceeds in two phases , as was shown in
Fig. 1.3. In that diagram, stage 5A of the system dynamics approach in-
volves experimenting with the model to see what happens when at least
some of the parameters and structures are changed. Parameter changes are
usually quite large, such as halving or doubling existing values. The purpose
is to deepen understanding about why the model behaves as it does, to
discover the range of behaviour of which it is capable and to stimulate
imagination about the structural changes which might be possible. With
that deeper understanding, analysis can move on, in stage 5B, to use optimi-
zation methods, which, guided by measures of what the system is supposed
to be achieving, search the parameter and structural possibilities much
more thoroughly than one could hope to do by 'suck it and see ' experimen-
tation. That is the real design phase of the system dynamics method'. This
chapter concentrates on policy experimentation; Chapter 9 will deal with
optimization.
BUSINESS PROCESS RE-ENGINEERING
The process of policy analysis in system dynamics models involves, as we

shall see in the DMC model, analysing flows of information in the model to
see how the actions thus produced generate consequences in the behaviour
of the system. The information/action/consequences paradigm discussed in
Chapter 1 is brought into playas an analytical concept as well as the basis
on which models can be constructed and tested. This process of experiment-
ing with and designing the flows to produce good dynamic behaviour is not
new, but is the essence of what has recently become popular as business
process re-engineering.
THE DOMESTIC MANUFACTURING COMPANY
The problems of the domestic manufacturing company (DMC) were en-

countered earlier and this will be the first case to be studied. Although the
model is fairly simple, it contains the seeds of so many solutions to DMC's
I The author realizes that some of his system dynamics colleagues, whose views he respects,
will differ from this interpretation. They will argue that the feedback loops give all the
guidance one needs to design good behaviour into a system. The difficulty is that a model of
even moderate size can contain many thousands of feedback loops.
problems that it is a good training ground in letting the imagination loose in

a disciplined way to explore a system.
For convenience, the influence diagram from Fig. 2.7 is redrawn in Fig.
7.1. The layout of the diagram has been changed somewhat to allow
space for additional influence links which will be introduced as we study
experimental possibilities. Figure 7.2 reproduces, for ease of reference, the
behaviour for the base case and with smoother responses , which was shown
in Figs. 6.3B and C. The extremely unstable behaviour mode in the base
case is credible enough for our purposes, and we have already decided that
it would be a waste of effort to complicate the model to represent
catastrophic behaviour more realistically. The model is on the disk as
FIG7-2.COS.
The Raw Materials Manager's pet theory
The Raw Materials Manager is heavily criticized within the firm for the
extreme volatility in Raw Material Stocks, and by DMC's suppliers for the
great variations in Raw Material Order Rate. His defence is that the large
number of different raw materials used in making washing machines, and
the inadequate clerical and computer support which he has, make it impos-
sible to keep even reasonable track of what has been ordered but not yet
received . It is, therefore, impossible for him to avoid the over-ordering and
excessive cutbacks which underlie the extreme swings in the raw material
variables. In short, his view is that the company's troubles are due to its
failings, not to his incompetence. His solution, or pet theory , is that he
should be allowed to invest a considerable sum of money in acquiring,
and subsequently operating, a Management Information System (MIS) to
allow him to control the quantity of raw materials in the delivery delay
pipeline.
Our first task is to find out whether or not it is true that the Raw Materials
Manager's 'pet theory' will bring the system under control. Testing the
effects of putative information systems in the inexpensive world of the
model is a most fruitful area of system dynamics application. Of course ,
we have already seen in Fig. 7.2B that there is a solution to the problem
of achieving stability which involves no more than changing the policies
which control the system, and that has the attraction that it would
cost nothing to implement. If the MIS also achieves stability, the question
will be whether it is 'better' than the policy solution and, if so, whether the
improvement is worth the cost. There might, of course , be other free solu-
tions which are better still, and we shall search for them when we have
tested the MIS.
The first stage of testing this pet theory is to change the influence diagram
to represent the new system. This requires us to model the contents of the
pipeline between RMOR and RMAR. This is the actual Raw Material on
A:.::c' ~0I1
(WEEIO Average
--~ ~~
W-b 01 A.v.mge
:~M~):EEK) PtoducNon in
De.u.dRaw
/ " I I ,/ ...... Ala_rial Sfoclc.
/ I I ;' WAP
RowDes~~11Q1 ..""" (Wf(K)
/ I \ Stocks
_ ..... DRMS <,
// , \ 1~;::3uc~,;:;{. ;' + (WMS)
/ I \ ~ ,, ;' \
\
/ I \ \ "
/ I \ \ / " \
I I \ \ / ;' ------ \
-t"' _ . . . .
I I \ \ / / .>
\ ..... <,
I I \ \ / / I ,
+y \ /
I \ ~ Average /
I Des ned PrOd uc fion
Bac klog , Rat e
,
I
'" \
I °b?t'~.. <, \ (WM~~EEK) I \
I (WMS) ' \ "
I 11 -, \ ,+ I \
\
I / \ \ I t+ \
, / -~ \ , Raw Material \
_ ~ Sf OCk ~eponcy ........ \
I / Back log \ I ;'
I =c:.AO:: + DoscE~ncv \ I / (WMS) <, +\
Iocklog / '" (WMS) <, \ I /
I WAO / \\ I ... \at, Material
I (WUK) / \ \ I Row Mole,,,,' Orde r Rate
I \ \ ' _ S~kS - .. (W~~~~EK)
I I \ \ '~(wMS) nm./o COINCI /'
'QwlrfaI.noI / '
1 I \" \ ~ Di"~r
<, _ . . (WEEK)
I, \
~'" Proocc, non
\ + Rat e
I \ ~WMtt~EEK)
, \ <41 R~r':'N~~~1 +
I Bac klog 0' - / RMAR" _
\ + ~~ / (W MS/W EEK)
\ .--" (WMS) /
/
r,.,.. Jo Correel
Ne w Order 0,., Backlog
Innow r~ote Dacl.pancy
NO ~IL
(WMS/wEEK) (WEEIO
Fig. 7.1 The original influence diagram for the DM C problem.

250
225
200
50
25
TIME (MEEK) SIMlA.ATED TINE

(_/WEEK) NEW ORDER RATE
(WNS/WEEK) PRODUCTION RATE
(_/WEEK) RAW MATERIAL 0RllER RATE
2000
500
250iI
I
o·
TIME (MEEK) SIIlIA.ATED TIME
(_ ) ORDER BACKLOG
(_) OESIRED SAClCL08
( _) RAW .... TERIAL STOCKS
(_ ) OESIRED RAW .... TERIAL STOCKS
A - RESPONSE TO SINE WAVE

Fig. 7.2 The domestic manufacturing company.
260
225
200
176
160
125
100
76
50
25
11 2 e
N
~·~MOR
TIME (llEEXl SIMl1.ATED TIME
/ llEEX) HEll ORDER RATE
IlIEE1CI PRODUCTION RATE
/MEEK) RAW MATERIAL llAllER RATE
2000
.
iL'~LOG
!III
~:~a~OG
TIME (WEEK) SIMl1.ATED TIME
I-I ORDER 1lAClCL08
t_ D1!&IRED BACI<l08
I_I RAW MATERIAL STOCI<I
l-.·-ORMS I_I DESIRED RAW MATERIAL STOCKS
B - LOWER STOCKS AND MORE SMOOTHING
Fig. 7.2 Continued

Order, RMOO, and, if it is required to control that variable, we must

introduce a negative feedback mechanism to do so.
As was pointed out in Chapter 1, negative feedback requires an actual
state, a desired state , a comparison between the two and a policy for
reacting to discrepancies. To meet those requirements, we must introduce a
new variable for the Desired Raw Material on Order, another for the
discrepancy between desired and actual raw material on order and a time
constant for correcting that discrepancy. This will give the simplest form of
control: proportional control. It is important to grasp that we must intro-
duce these factors because they are required by the laws of negative feed-
back, which must be satisfied if control is to be achieved.
The laws of negative feedback lay down what must be achieved, but not
how it is to be done. Thus, the designer of managed systems has complete
liberty to model factors such as the Desired Raw Material on Order,
DRMOO, in whatever way gives the best performance for the system. It is
this degree of freedom which makes designing control policies and struc-
tures for managed systems a practical possibility and an interesting task. By
analogy with Desired Raw Material Stock, DRMS , which is also a target
condition for a negative feedback loop, we shall initially base DRMOO on
the Average Production Rate, which is the rate at which raw materials are
used up, multiplied by the number of weeks of raw material usage which it
seems plausible to have in the raw materials pipeline.
These influences are shown on the right-hand side of Fig. 7.3, the new
links being emphasized by thicker lines. The reader should carefully study
that diagram before proceeding. For instance, have solid and dotted lines
been correctly used? Are the signs correct ? Notice the new notation for
Switch 1, with an arrow pointing to a solid circle on the link from EPS3 to
RMOR. This is to indicate that we may wish to run the model with and
without the MIS, the latter being the current system. If the MIS is turned off
by having Switch 1 set to zero, the new variables will still be calculated
within the model, but they will have no effect. Switch 1 is, therefore, a
structural parameter.
The equations required to model these changes are fairly straightfor-
ward.
r rmor.kl = eps2.k/tarms + apr.k +switchl *eps3.k/tcrmp
c tcrmp=4
The order rate equation acquires a new term , activated or suppressed by
Switch 1.
note
note new sector for management information system
oo~ ==========================
note
note switch to test effects of MIS
~~""'.t.
oI.towWaf. . .
/
"=..'"
D.....
----- / (WEE«)
,/
..- - ........
A:=~ ,/ <, /
(WfEIO
,/
-- <,
Av..o /
.."" O rd erge
/
.......
.........+ ,.
/
- +1WMS/W"
- .. ~6:: 1O / ... Do uedRow
Mo 1eriol on
.> ,I / w..n 01Av.ra"...
De seed _ _ o...d law D~ -..
/
//' "
I
\
\
A.__
... _ ..,...- I __
_ __ -- _
Raw Molena! .....-
Stocka
- . DR' ... -
-
_
MaMIIaI:IIocb
...WA~
_ ...-
(WMS) "
+ ~
(W MS) " '\ \
' " I / \ \
/ /
\
/' '
" \
~,\ I / __ ----- \ I
- t-,
" " \ , ,
/' '\ \ 1,
/"/ <,
I
I ' \ 11 Av..o ge " \
' ",od uc''''" I " -, -, I
' "" lI: I
I '' 010''''''' "
= ,\ I I \ I
\
-~'
-, " I \ I
0;::;:; " I / ~+ \
/
'' ",
Raw Mat eria l \
' .
, -,, ,
,, , _ ....,..stoc k ~epon cy
(WMS )
_
\
Switch J
I
t+
Row Mo to llOJ
on Odel
' ", , " '\ +\ -- l>tIc leponcy
.i-- 'PSJ
+ \cl Ma tenol ,y (W MS)
:
I
"
... 0 / \ -, ,
, '"="' Q rde, Ro fe - A-
........... ~ (W~~EK) ,
(UR) \ -, ,
: =.l.=
I I
;/'~
- """ \, '\ '\
: ~_ (~ ) 1m. 10 eon.t:t
.,..".".tId .....
I
I , -, " "':'_ -- " " -, I
(UEI)
-~
' I ;_ 'Jl.O __ <, "::w-:.ac::::' I
I \ A"1lMAA
r a.. I
~MS/W"K) (UEIO
=.~ _ -
, • +
/
IlLOG /
,\ +
(W MS) I
/
'' II ~"
" -- - _~C~ ,
._~" 0_ .
_=- /
\ 0<d0,1oc_ /
,/
K) --~."
,,"w o.de' ..-
Inflo~Ole
(WrvtS/WEEK)
I _. ' Ali
(Wfl1O
-
Fig.7.3 Th e Raw Mat erials Man ager 's information system in the DM C probl em.
note
c switch1=1
note raw material on order
note
I rmoo.k = rmoo.j + dt*(rmor .jk- rmar .jk)
n rmoo=drmoo
note
note desired raw material on order
note
a drmoo .k=apr.k*dwrmp
n dwrmp=ddel
note
note pipeline discrepancy
note
a eps3.k=drmoo.k-rmoo.k
The equations for the raw materials pipeline and its control are exactly like
those for, say, BLOG. Note the use of
n dwrmp=ddel
It seems sensible for DWRMP to be the same as the delivery delay, and this
use of the N label sets up a 'computed constant' in which one constant is
determined from another. This prevents errors such as changing DDEL, to
represent a change to suppliers who can deliver more quickly, but forgetting
to change the numerical value of DWRMP.
d drmoo=(wms) desired raw material on order
d dwrmp=(week) desired number of weeks of raw material usage in
x the raw material pipeline
d eps3=(wms) difference between desired raw material on order and
x actual raw material on order
d switch1= (1) switch to test effects of introducing the raw materials
x manager's management information system
d tcrmp=(week) adjustment time on raw material in the pipeline
The amended model is on the disk as FIG7-4.COS.
The effects of implementing this fairly costly MIS are shown in Fig. 7.4.
The upper graph shows that the MIS does, indeed, bring the system under
control, in the sense that the explosive behaviour of Fig. 7.2A has been
replaced by sustained oscillations, but the behaviour of RMOR is worse
than that in Fig. 7.2B. The control of RMS is poorer and stock levels are
higher because WAP has not been changed from the base case.
When TCRMS is increased (Fig. 7.4B), the behaviour of the system with
the MIS becomes worse. The oscillations in RMOR are of greater ampli-
tude , as are those of RMS. This difference of behaviour between the two
The dom estic manufacturing company 205
250
225
200
17 5
15 0
125
50
25
TIME (MEEK) SIIU.ATED TIllE
(MilS/MEEK) NEM ORDER RATE

(MNS/lfEEK) PAOIlUCTION RATE
(NNS/MEEK) RAM IIATERIAL CROER RATE
2000
TIME (MEEK) SIMULATEO TIllE
(MNS) OAOEA BACKLOG

(MNS) OESIRED BACKLOG
(MNS) RAil MATERIAL STOCKS
lNNS) OESIRED RAM MATERIAL STlICICS
A - RESPONSE TO SINE WAVE

Fig. 7.4 The management information system.
250 ,
225
200
175
150
125
100
75
50
25
I L' TI ME lWEEK) SIMlA.ATED TIWE
1--~2
lWWS/WEEK) NEW ORDER RATE
lWWS/WEEK) PROOUCTION RATE
~'RMOR lwws/WEEK) RAW WATERIAL OROER RATE
2000
TIME (WEEK) SIWlA.ATED TIllE
1_) ORDER BACKLOG

lWWS) DESIRED BACKLOG
lWWS) RAW IlATERIAL STllCKS
lWWS) DESIRED RAW IlATERIAL STllCKS
B - MIS WITH SLOWER REACTIONS
Fig. 7.4 Continued

halves of Fig. 7.4 drives home the essential insight that simply implementing
an MIS is unlikely to be effective in any system, not just DMC, unless one
carefully studies the system 's policies. Introducing the MIS has changed the
system, so it is intuitive that the old policies might not be good for the
changed system.
We conclude from this that implementing the MIS, with the existing
policies, is by no means as good as changing the policies. Of course, it is
possible that implementing the MIS and changing the policies might pro-
duce even better behaviour.
Parameter analysis
The range of behaviour already found suggests that some more experiments
are called for, but it might be as well to be a little more systematic than
simply trying a few at random, as we did in Fig. 7.2B. For these experiments
we shall switch off the MIS and not consider its parameters.
Apart from the parameters which specify the behaviour of NO, the DMC
model has seven constants:
• WAO: the number of weeks in advance which the customers place their
orders. This would be difficult to change without upsetting them.
• DDEL: the delivery delay on raw materials. Again, hard to change
without renegotiating the whole system of raw material contracts.
• WAP: the number of weeks of average production to be held in Raw
Material Stock.
• TAOR, TAPR: the two averaging times for New Orders and Production
Rate, respectively.
• TABL, TARMS: the two adjustment times for Backlog and Raw Ma-
terial discrepancies, respectively.
The last five parameters are entirely at the choice of DMC's managers and
can be changed freely. They are , however, of two distinct types.
• WAP acts as what is called a gain in control engineering. That is, it acts
to multiply the variations in APR when deriving DRMS. WAO is also a
gain.
• TAOR, TAPR, are delays involving the smoothing of information, as
was discussed in Chapter 4, p. 105.
• TARMS and TABL are parameters in control equations and, as ex-
plained on p. 102, they also act as delays.
It is a rule of thumb in control engineering that reducing gains and increas-
ing delays is likely to reduce instability. When that rule is applied to the
gains and delays in the model, using FIG7-4.COS on the disk, Fig. 7.5 is the
result.
In Fig. 7.5A, the stability of RMOR is very slightly better than Fig. 7.2B,
showing that the extra parameters, TABL and TAOR, have had a practi-
TIME (MEEK! SIMULATED TIME
(MMS!WEEK) NEM ORDER RATE

(MMS/MEEK) PROOUCTIOM RATE
(MNS/MEEK! RAM MATERIAL ORDER RATE
2000
TIME (MEEK! SIMLl.ATED TINE
(MNS) ORDER BACKLOG

(MNS) DESIRED BACKLOG
(MNS) RAW MATERIAL STOCKS
(MNS) DESIRED RAM MATERIAL STOCKS
A - DOUBLE AND HALVE GAINS AND DELAYS
Fig.7.S Parameter sensitivity for DMC.

250
225
200
175
150
125
100
75
50
25
It
IL_~~
TI ME (wEEK) SINUUTEO TINE
(WHS/WEEKl NEW ORDER RATE

(WHS/WEEKl PROOUCTION RATE
~ -RMOR (WHS/WEEJ<) RAW NATERIAL ORDER RATE
2000
TIME (WEEKI SIHU..ATEO TINE
'I'll--OBLOG
L 'BLOG (WHS)
(WNS)
ORDER BACKLOG
OESIRED BACKLOG
'~'RMS (WHS) RAW .... TERIAL STOCKS
L..·-DRMS (WHS) OESIREe RAw .... TERIAL STOCKS
B - MIS AND GAIN /DELAY CHANGES
Fig. 7.5 Continued

cally negligible effect. That is not surprising, as those parameters relate to

the production system and there is no feedback relationship between the
production and raw materials departments.
The conclusion we draw from Fig. 7.5A is that the policy solution to the
problem of bringing the system under control is spectacularly better than
the MIS solution, shown in Fig. 7.4A. Since it is also a free solution , there
seems to be no contest. It is, of course, possible that the costly MIS com-
bined with the free policy solution will produce even better behaviour, and
the test of that is shown in Fig. 7.5B.
The addition of the MIS, in Fig. 7.5B, improves the stability of RMOR
quite significantly, and produces a slightly better match between RMS and
DRMS . In both Figs. 7.5A and 7.5B, the match between BLOG and
DBLOG is virtually perfect. Stability of RMOR is nice to have from the
suppliers' point of view and probably eases DMC's cash flow a little, but the
truly significant variables for DMC are the ability to match BLOG to
DBLOG and to keep RMS close to DRMS. The performance in Fig. 7.5B
is so little improved in these respects that the MIS seems to be hardly
worthwhile.
Structural analysis
The preceding experiments with parameters only show that quite consider-
able improvements can be made to the quality of the system's performance,
and this encourages us to become more wide-ranging in our thinking.
Consider , for instance, the equation for Raw Material Order Rate ,
RMOR:
r rmor.kl=eps2.k/tarms+apr.k
The APR term exists in this model of the Raw Materials Manager's de-
cision-making because he wishes to ensure that he continues to match the
current level of usage, as well as correcting stock discrepancies. We
know, however, that the delivery delay on raw materials is 6 weeks, which
happens to be exactly how far ahead the customers order their washing
machines. It might, therefore, make sense for him to use AOR in his
equation so that he orders raw materials now for delivery 6 weeks hence,
which is just about when the orders now being received will be in pro-
duction . He does not do so, because the poor personal relations between
him and the Production Manager mean that order rate information is not
passed on. This may seem crazy, but is exactly the kind of thing which
happens in real systems and which can considerably complicate the task of
the analyst.
If, however, some more senior manager directed that the Raw Materials
2 At the risk of repetition, this case study is adapted, and somewhat simplified, from a
consultancy project. The company really did have the policies and probl ems described.
Manager should have access to order rate information he could use the
equation:
r rmor.kl=eps2.k/tarms+aor.k
Intuitively, one would expect this to give an improved performance, but it
might be that a mixed policy would be even more satisfactory. Replacing
the existing equation for RMOR with:
r rmor.kl=eps2.k/tarms+(alpha*apr.k+(1-alpha)*aor.k)
will allow us to experiment conveniently with different options.
ALPHA is a weighting factor between two streams of information, and is
shown in Fig. 7.6. It is a structural parameter because, if the best perform-
ance is found when ALPHA=O, the dotted link from APR to RMOR could
be deleted and the thicker solid link from AOR to RMOR would be
retained, thus changing the structure. That would also imply that the firm's
management information structure had been fundamentally wrong; in the
real world, the Raw Materials Manager would have to switch to basing
order rate decisions on AOR and abandon APR, except for determining
DRMS.
If, on the other hand, the best performance is achieved when ALPHA =
0.5, then both links are required and, in the real world, the Raw Materials
Manager should use both information streams, giving them equal weight in
his decisions.
Of course , Switch 1 is also a structural parameter, but it can only have the
values of 1 and 0; the management information system which is the Raw
Materials Manager's 'pet theory' either exists or it does not. Switch 1 is said
to be a binary structural parameter. ALPHA, on the other hand , is a
continuous structural parameter, which can have any value from 0 to 1,
inclusive. Naturally, the combinations of ALPHA and SWITCH1 are all
valid: ALPHA not being 1 does not rule out SWITCH1 being set to 1 to
represent the combination of the changed information flows and the MIS3.
Testing the model with the changed equation for RMOR (see FIG7-
7.COS on the disk) produces the behaviour shown in Fig. 7.7. Because using
AOR in the RMOR equation is intuitively appealing, ALPHA is set to 0
and the MIS is turned off in Fig. 7.7A and on in Fig. 7.7B.
Comparing Fig. 7.7A with 7.5A, the structural change involving ALPHA
has had very little effect. The oscillations in RMOR are about as large, but
they are now slightly closer to matching those in PRo There has been no
detectable effect on the behaviour of RMS and BLOG compared with their
respective targets.
3 As we remark ed earlier , the system dynamics approach to evaluating the structure of a

management information system before one considers hardware and software solutions has
proved to be very fruitful.
o..JfWldWNb
oIl1aw~1kIfa
~ ...
--.------- / 't::~
--- -- ........
-~-- /
A:,==" _ A, ..oge_ - - - ....... --
/
fA"" <,
/WEEI() , m _ .. O"'e ' /'
/ ' ' ' ' ' - /
....... + ,.
.> .... / " "
I /
// "
~,. I\ <,
_""'A__ "- '4.~~~ O roer -..
/ I \ / "
A / DeSl'OO .1(' / '\ D~ -,
/ I \
.-<~~
~~.
/ I \ .. ~- Row Mote'''' /WE&:> "
, ~- ' \
" : " ~ I .> . > .. \
, "
-=-', ' \
\
I " \ , ,I / / --------' U-- --_ I
I " , " - > ---' , \
I Y \ \ 1"_ I I
De..ec \ '"
I / / > ,," \ -- \
Bockbg .. A,.. " <, ..f
~ J
I __ \ P<oduc~n \ -, \
,I I
o~_ ~. lJ::
, ! ' ,-,', '''''l''"'. ', -,' \
' " , \ I
/ " \
' \"
t+
,, -:
::::::.= " I, ' I Switch 1 Raw Motenol
, ,~-, , , I on o.de,
, - " _" " - > -- Decreccrcv
,
- /
' "", \ \ I
,-
-_.,
-~-~~'
~ ---- , <,
I
.4.- EP'33
(WMS )
, ' , , ' I -, , " - A-
, I , , ' -- '... ,
I I -, -, " -,
I, / - (WMS)
St~ " ..
R&",,,.,,:,&:.:"
RMOR
~ I
I \ <, ~ ,."... e-.c, ,," (WMS/WH K) ~ <,
, , - .+" , -- -- . . ., J
"- '" nm. Ib Con.cl I
I \ ~+P,O~ehon -.:::c' "' .,.,AoIcDIttIG
raw /
, ' ....%.~ - /WEEI()
-A BlOG / Row Mo,..."
\ ___ (WMS) • AnNal Rot e + DDEL /
\ +
Boc~cl
O'd"'- / / (W'=~EK) .. --
-- /
/
NO mo .. C.u.c/
/'
_ onO.der
Row Mo'e, ie'· +
n /WEEI()
/ ~ RMOO
(WMS ) -
~;:;'w°,;: I
(WMS/WH K) '='.. = ~ --
Fig.7.6 A structural option for the DMC problem.
250
225
200
175
150
125
100
75
50
25
it
IL-~2
TIME (wEEK) SIMlA.ATED TIME
(WMS/WEEK) NEW OFUlER RATE

(wMS/WEEKI PRODUCTION RATE
~ 'RMOR (OMS/WEEK) RAW MATERIAL ORDER RATE
2000
. 120 .
IIIL' TIME (WEEK) SI......ATED TIME
,'I: BlOG
--OBlOG
' ~ ' RMS
(WMS)
(WMS)
ORDER BACKLOG
OESIREO BACKLOG
(WMS) RAW MATERIAL STOCKS
l..-·-DRMS (WMS) OESIRED RAW MATERIAL STOCKS
A - ALPHA=O AND NO MIS
Fig.7.7 Output from structural option for the DMC problem .

250
225
200
175
150
125
100
75
50
25
It
IL-~2
TIME (WEEK) SIN\A.ATEO TINE
(WNS/WEEK) NEW ORDER RATE

(WNS/WEEK) PROOUCTION RATE
- -RMOR (WNS/WEEK) RAW NATERIAL OROER RATE
2000
1750
TI ME (WEEK) SIMLLATED TIME
(WNS) ORDER BACKLOG

(WMS) DESIRED BACKLOG
(WWS) RAW MATERIAL STOCKS
(WNS) DESIRED RAW MATERIAL STOCKS
B - ALPHA=O WI TH MI S
Fig.7.7 Continued
When Fig. 7.7B is compared with 7.5B, the addition of the MIS has
brought RMOR into practically perfect agreement with PR , though there
is, again, no visually detectable difference between the behaviour of RMS
versus DRMS and BLOG versus DBLOG between the two graphs. Once
more, the MIS seems to be hardly worth the cost.
Just because ALPHA seems to have had little effect on the system does
not invalidate the idea of changing structures, and it is worth looking
imaginatively at Fig. 7.6 to see what further changes suggest themselves. A
collection of possible changes is shown in Fig. 7.8, although that does not
exhaust all the options. For instance, the EPS1 and AOR policy for RMOR,
mentioned on p. 196, is not shown.
The changes involving the continuous structural parameters GAMMA
and DELTA are very similar to that involving ALPHA. In control engi-
neering terms, all three of these parameters involve feedforward. The effect
of the external driving force, NO, as measured by AOR, is transmitted
directly forward into the system. The corresponding effects on the
equations for DRMS and DRMOO are:
a drms.k=wap*(gamma*apr.k+(1-gamma)*aor.k)
and
a drmoo .k=dwrmp*(delta*apr.k+(1-delta)*aor.k)
Obviously, if SWITCH1=O, the value of DELTA has no effect on behav-
iour, as DRMOO is effectively neutralized.
Note that in FIG7-9.COS (and in FIG7 -7.COS), the equations involve,
for example,
a drms.k=wap*(gamma*apr.k + (one - gamma)*aor.k)
c one=1
d one=(1) the number 1
This is good practice for any system dynamics software package which
supports dimensional analysis (some do not) , as it defines the number 1
on the same basis as GAMMA. This allows the dimensional analysis soft-
ware to make a valid comparison between the two to ensure a
thorough check.
Assuming that Switch 1 and Switch 2, which will be described in a mo-
ment, are both zero, the parameter BET A will introduce a third feedback
loop into the system. It has the effect of linking PR to its own average ,
thereby introducing an element of damping into the system. The effect of
this should be to reduce the variations in PR , on the grounds that changing
production rates is usually a very expensive business.
Switch 2 has the effect of modifying PR according to the availability of
raw materials. If EPS2 is positive, raw materials are in fairly short supply,
so it might be a good idea to ease back on production to prevent the
problem getting any worse. If EPS2 is negative, raw materials are plentiful,
".,..----- ----- --- d __
.,/ 0.,., .......
.,/
---;;r-- . . . . ~~"..
»: ~" <, <,
/' . / .--- ----- "- "- /=

--....- --- z-:« --- /
"==- /' - .,/ -- ~ ........ /
,., /' ./ - »: ................
TAc.~ f1tUIO - _ _ Av8l'og&./ /' .........
-- ........
" "- ..
l'AQR
- .. ~- - - - - - - -- --
/ ........ ,
................ 'c "./
/~IWEEI(} 7......... .......dA....~ ......
,,--.. ':...~ta%R~
O<de<
/ DIlMOO
/
/" I" , / /
/..............
.... Deued .....
//-
"""
-
=""..;' '" "-
MMS) -,
I I \ / I Raw Mat_ fWfBO "-
I , \ ~--=. y --'g.,~ <, -,
/ 1\ MI'R I \ / ... - l'CWMS) '\ \
/ f \ ~ I. - : \ \ \
/ ,\ , ' , / \ (X- - - - - - -
I f \ ' I /" ,,- - - - - - - - - _\ I \ I
I I \ , / /" 1"- I \
I I \ ' I / // \ - J <, I
I .. " \ \ I / \..... \
I DeUod \ "-,~/ \ -, I
J :X~ \ Pr~. I " \
O6lOG - \ /. AP1l
I lWMi) ..... -, / (WMS/WEEK) I ' I
I " -.", f '
, / \ ,,-" P-., IA.. Switch 2 ~ ',
,,
'I ..
,, I I -
Ilodolog
~, , I 'I II _ -- _ .ow....,.:...
- - -_stoe. Ooc,epancy - - _ \ / Switch I
I ........G/ Awweve ObcrfljlOnCy \ , ......... ./ - EPS2 <, -, \ I RON Materiel
OrdIN_o.tIMd +.,. EPSl '\ \ I Y /' / ('NMS) -, \ +I .. on Odel
I "'::".... / (WMS) ' " I • I .. ......, .. .. -- -- DIoCfi,'i;'''''Y
I _SO / " \ I / I ]«,;"Mat""", . - - - <WMS)
I / " I / -~- A-
I I -, \ - OMS .» lWMS/V,{EIQ
~e ~ I
I \" J / ~ /WMS) n.-_,..~ ..... - r :,
I I " , \.. , - "=:=c:/ / \ <, "
I \ <, .. ,,~ f rI _______--
_ _ .t.. u_
\ iWiiKJ / \ <, n;;"*:t::::'
\\
\ \
~;-"P'T--- lWMi/WEEK) \ , / \ '= fWfBO
\ .. RowMot ' t
\I .. 8oc'log
0<_
o' _ /
/ - , "''''''''.::::;'
RMAA
..
~LJDDEl ~) /
\ BlOG / MMS/WE£IQ
New~de' (WMS) '::::~ \ Raw Materio' .. ,///

'"""NORa,. ~~ »>:
<WMSlWEEJ<) ~ (VvMS) - - -
_r
Fig.7.8 A collection of structural options for the DMC problem .
so it might be worth taking advantage of the surplus to reduce backlog so

as to be in a better position to cope with any surge in orders which might
happen.
BETA and SWITCH2 alter the production rate equation to:
r pr.kl=epsl.k/tabl+(beta*aor.k+(1-beta)*apr.k)
x -switch2*eps2.k/tarms
Note the minus sign before SWITCH2. If EPS2 is positive, we need to
reduce production.
Strictly speaking, we should not use TARMS in the above equation.
TARMS represents the time constant in adjusting raw material stocks by
ordering new supplies. Here we are using raw material stock to control PR ,
and there is no reason why the same time constant should be used. The
reader should alter FIG7-9.COS to use a new parameter TARMS2 in the
equation for PRo This has the advantage of creating a fresh policy analysis
option, as it may be that different values for TARMS and TARMS2 pro-
duce the most effective behaviour for the system.
Switch 2 is potentially a very important feature because of the effect it has
on the system's feedback loop structure. The original loop 1, shown in
Fig. 2.7, controls production, but loop 2, and the new loops which are
created if SWITCH1 is 1, operate only on the raw materials sector. Loop 1
drives loop 2 through PR but, without Switch 2, there is no feedback from
raw materials to production. Switch 2 therefore has the potential to create
an overall linkage tying the two parts of the system together. Since the new
loop is negative, it ought, in principle, to improve the ability to control the
system.
What values seem plausible for these new parameters? Let us keep
ALPHA to 0 on the grounds that raw material ordering should be geared to
the level of business. We set BETA to 0.5, because it seems a good idea to
strike an even balance between PR being flexible to meet changes in AOR
and damped to avoid unnecessary changes. GAMMA=0.5 also seems a
plausible value, though the reader should work out why. We shall leave
SWITCH1 and SWITCH2 at 0, so as to allow you the chance to make your
own experiments.
The dynamics with these values appear in Fig. 7.9, and we shall need to
compare that with Fig. 7.5, which was, on balance , the best option found so
far.
In case A, without the MIS, the structural changes shown in Fig. 7.9 are
a general improvement over Fig. 7.5. The oscillations in RMOR are a little
smaller and come practically into phase with those of PRo However, as we
have already mentioned, this improvement is nice to have, but is not the
crux of the matter, which is to control BLOG and RMS to their respective
targets. In those respects, Fig. 7.9 shows a slight improvement in the RMS/
DRMS comparison at the expense of a slight deterioration in BLOG vs.
DBLOG.
200
175
150
125
100 -+-t+-_......
75
50
25
TI ME (lfEEK) SIMlA-ATED TIME

(lfMS/lfEEK) NElf ORDER RATE
(lfNS/lfEEK) PRODUCTION RATE
(lfMS/lfEEK) RAil lfATERIAL ORDER RATE
IIIiL 'SLOG
TIME (lfEEK) SIMULATED TINE
I. ....::.~~~§OG
(_) ORDER BACKLOG
(lflfS) DESIRED BACKLOG
(lfMS) RAlf lfATERIAL STOCKS
L.- ·-ORMS (lfNS) DESIRED RAil lfATERIAL STOCKS
A - BETA. GAMMA = 0.5. ALPHA = 0 WITHOUT MIS

Fig. 7.9 Output from other structural options for the DMC problem.
250
225
200
175
150
125
1oo+t+-_...,.
75
50
25
O.........-+.---In.- :-~o . ao . 50. Bo . '0 . 130. 90 . 1100 . ~1O. ~20 .

TIME (WEEK) SIMULATED TIME
(MMS/MEEK) HEM ORDER RATE

(MMS/MEEK) PROOUCTION RATE
(MMS/MEEK) RAM MATERIAL OAllER RATE
2000
500
250 iI
I
o.I"""",,,----1.,..- 70. 130. 90 . 1100 . 1110. ~20 .
II J'Bb~8G
TIME (MEEK) SIMULATEO TIME
(MMS) OROER BACKLOG

(MMS) DESIRED BACKLOG
L:::§~~s (MMS)
INNS)
RAM I4ATERIAL STOCKS
DESIRED RAM .A TERIAL STOCKS
B - BETA. GAMMA 0.5. ALPHA o WITH MIS
Fig.7.9 Continued
With the MIS, case B, the structural changes in Fig. 7.9 leave RMOR
practically indistinguishable from PR, and the control of RMS becomes
much improved, though there is a slightly greater deterioration in the
control of BLOG to its target of DBLOG. Whether the improved control of
RMS is worth the extra cost of the MIS is a question we cannot resolve. To
some extent, careful costing studies might show the answer , though it would
be hard to find convincing cost values for the slight degradation in BLOG
control.
Why, though , is there such a perfect agreement between PR and RMOR?
The reason is not far to seek. Basing RMOR on AOR instead of PR, and
ensuring that RMS does not go wildly out of control by keeping track of
the quantity within the delivery pipeline ought to give good control, and
does.
It should now be evident that there are a large number of solutions to the
problem of keeping DMC under control. Some are only fairly good, such as
Fig. 7.4, which showed the effects of simply buying an MIS and plugging it
in. All the others are much better than that. Reducing gains and increasing
delays (Fig. 7.5) gives a very good result , to which some of the structural
options add further improvements. The MIS added to the parameter and
structural options seems to add relatively little.
There are , however, other structural options in Fig. 7.8 which have not
yet been tried , and Fig. 7.10 shows yet one more , which is to damp RMOR
by taking into account its own average, ARMOR. If that is to be considered
in combination with the possible use of APR and/or AOR, an additional
parameter, KAPPA, is need ed. One way of writing the equation for
DRMOO would be:
a drmoo.k=dwrmp*(kappa*armor.k+(l-kappa)*
x (delta*apr.k+(l-delta)*aor.k)))
Summary of DMC's problems
The reader should now see that a system dynamics model can contain very
rich policy analysis options. We have used DMC to illustrate this phenom-
enon, but that model is by no means unique. One of the skills of building
system dynamics models is to recognize the policy options they contain and
to be able to adapt the model to test the options which have been detected.
This is a skill which comes with practice, so the reader is urged to experi-
ment further with this model, trying some of the variations which have been
suggested, and others which can be discovered for oneself, such as adding
noise to NO.
It will also be evident that simply trying parameter and structural options
like this is rather a hit-and-miss process. One is using visual judgement of
the output to assess whether or not matters have improved, and one can get
trapped into an almost endless series of experiments in the hope that
something will be found . Paradoxically, this is an essential step on the way
- - - - - - - - - ---
o.•.,w....r.
- - oIlaw-=1l
<,
/ -- !=
/' / ~
f"; ?" - ~-~....
/' .......
/' -- - - - - - - --- \\ - - ........ <, /
A~~ /' --- ....... , <, ~" + /
rM>I / '/.. . ...... , - " \.,
MIlO
Av.._ //'
/'
/
-- --
- -
-- - - - - _ /
/'
--- - - ---~~ ~ .......
_.oIA_ <, ,
- - IC,
"
....... +
"'M '
,.
''''''''''edOOW
",
~
- - . Order - - ? _ _.. . . A__ <, DllMOO
" ....... ....... .... Malenclon
/ + - / / -/................ +
...... DOMed . '&-
- -=:::::::.. '\. =. .... \"'- _ +1f (W MS)
'"
'" / _ Mot",'" /WID) WNOO
~MS/WEEK)
~I I "" I
// I \ / / __ = _
, ,(WID)" '-'- " ' '\.
/ I \ Y // + (WMS) -, " Av",_ _ ' S wllcIJ J \
I / \
I
I
I \
A__
".~.. ....
I
/ \
\'"'" '"
/'
_ _____ , \
\ rv___ _ _
U;
I
\
\
- ' Ma l enal Ofd er
note \
I \ \ '" - _ _ I 11
I I'
M~ \ . r : r--_, . \ (w~~:EK)
+ I
f ,\ \
t / '" - :: \ " , /
f \ \ \ \ I
f ' \ \ I
1 r / ; '" "" /
/ +~ \ ":~/ I \ \ t
I """'ed \ Il<>t. I \ I
I --
oIc.~ " \\ / /'(W~"'' ' 4 IJ \\ II I
I \ A i" ' 2 l' + I I
: / ' \ -'I JI , s..,'fch "
~)" \~,/ ~~~'!""_
' / - II \ ' I Switch 1
...- _ -~"·"""'''' EPSr '''' J --- _....... \ J" I '+
/ oJ., \ : / / (WMS) '+_ + __ A_
'I __._ - ..,... , \ .•L, - / + __ .... I
' _+~.
•
,I / +,,,, _ 1 ..-
I a-. .. - +/ .. (WMS)"" , /' _ Mot..o /' .. (WMS/WEEK) " -
..... / \ \ I Stocke / ' /' I
'
I '''0 (WIll) I \\ ' , /
~ _ (WMS)
OMS __
.-..c _ /' /' "
I' I/ \" ' , , \ , ' I, -.:"'r

_ ' , . - ..c _ _.=- /I
I I , re-
+"''41' r / ~
' , -=- - I
I \ ~ O""Mot..." + /
~. (WMS~EE K)
I \ ~ A'ti:'A'jf'e DOH
'\ ,__ 0' - I ' _. ~ /
+ go~ f + /'
I (WMS) f _ Mate"", /
\ f _
NewOrd e, "'--loG
.-.. e--,
~If. (WMS)
Inflow (lft'1I)
'1!~ _
NORote
(WMS/WHK )
_r --
Fig.7.10 Ye t anothe r struc tura l opt ion for the DMC problem.
to deep understanding of a model, but it is also clear that there ought to be

a more comprehensive approach. Fortunately, the use of optimization,
which we shall study in the next two chapters, is such an approach. Before
we use it, however, we must experiment with another model.
PRESSURE POINTS IN THE COMBAT MODEL
The DMC model is a case in which one experiments with the parameters
and the structure in order to obtain good performance. The system is,
however, so heavily dominated by the external driving force, in that case the
incoming new orders, NO, that the overall mode of behaviour tends to
change in relatively fine detail, once the designer has cured the explosive
behaviour mode of the base case. There is, however, another class of model
in which the behaviour mode is determined entirely by the structure and
initial values, not by a continuous external driving force; the combat model
studied in Chapter 3 is such a case. In these undriven models, the behaviour
mode can change quite remarkably depending on variations in parameters
and initial values. These parameter changes usually correspond to invest-
ments which its managers might make in aspects of the system; we referred
to those aspects as 'pressure points' in Chapter 3.
Pressure points and policies

The concept of a pressure point is rather different from that of policy which
was discussed earlier in this chapter. We saw there that 'policy' meant a
guiding rule which would be applied continually as time passed and circum-
stances changed. It specified how actions should be taken in the informa-
tion/action/consequences paradigm, and there was an emphasis on the
repetitiveness of the actions . Pressure points, on the other hand, carry an
implication of an investment or a choice which is made only once, after
which the system will run under the influence of that choice. Naturally, one
seeks to find the pressure points which will make it run in the most effective
manner possible, so pressure point analysis is still policy analysis, albeit of
a slightly different type. Equally naturally, there is no hard-and-fast divid-
ing line between the two modes of analysis and, in practice, a given study
may involve both types.
The combat model is a good example because it is relatively simple and
contains numerous pressure points. This multiplicity of possible invest-
ments leads very naturally to the idea of the optimal investment when funds
are limited, which we shall explore in detail in Chapter 9. Military modelling
has been a very successful application area for system dynamics, so the
model is interesting in its own right though it is very simple and does not
begin to show the true subtlety of system dynamics models of military
problems. Of course, since it involves an interplay between two non-coop-
Pressure points in the combat model 223
erating sides, it has implications for the modelling of competition between
businesses and nation states.
Experiments on the pressure points

For convenience of reference, the influence diagram is reproduced in Fig.
7.11. Three of the pressure points are emphasized by thicker ellipses, as
they are the ones we shall test. The reader has several other options to
investigate, though the Information Collection point is for the advanced
student. FIG7-12.COS on the disk sets up experiments to produce the
graphs shown in the successive pages of Fig. 7.12 for our chosen pressure
points. It is also set up to print values of the variables at the beginning and
end of each run. This use of printed and plotted output is far better than
relying simply on graphs of a few variables. As a basis for compar ison, Fig.
7.12 A shows the base case in which Blue is defeated by TIME=33, Red
having 34140 troops left out of his original total initial strength of 65000.
Figure 7.12B shows the effects of increasing Blue's transport capacity
from 4000 men/hour to 6000, perhaps by buying more vehicles. This is a
fairly significant change, but pressure point testing is best done in large
steps initially to get directions of improvement. Blue's defeat is staved off
by about 2 hours, Red being left with 29800 survivors; a better performance
by Blue. The interesting feature of the lower half of Fig. 7.12B is that Blue
starts to commit reserves at TIME=12, as before , the 20000 men available
being just sufficient to restore the force ratio to the acceptable value.
Unfortunately, by TIME=24, Blue is again too heavily outnumbered, so
the Force Ratio Trigger goes back to 1, but there are no reserves left to
commit.
Figure 7.12B showed the same phenomenon as DMC: differences in
performance relative to Fig. 7.l2A, but within the same overall behaviour
mode. In Fig. 7.12C, Red Transport Capacity is reduced from 2500 men/
hour to 1500, perhaps by Blue sabotage teams. It is not assumed that this
change would cost the same as the previous one; it is just another experi-
ment. This change produces a change in behaviour mode. Blue now wins,
with nearly 5700 men surviving. The reason is that Red's reserves trickle
slowly onto the battlefield over a period of 33 hours and Blue's reserve
commitment at TIME= 12 is enough to give him an overwhelming advan-
tage from which Red can never recover.
Figure 7.12D is the result of the experiment of reducing Blue's movement
delay from 6 hours to 3, perhaps by buying surveillance equipment to locate
Red forces so that Blue's reserves can be positioned more effectively. Blue
now survives until TIME=37, and Red is left with 22800 at the end. The
behaviour of Blue's reserve commitments is interesting. Some are sent
forward at TIME=12, as before , but commitment stops a few hours later,
when there are still a few reserves left. Blue's commander needs to commit
reserves again at TIME=26, but there are too few to save the day. Nonethe-
.-----,
.-s;u. commande,',? )
Acc.plable Force
Ratio
<,
~ + -,
\ Force Pc no k'"'
\ __ - -- (, "J(:I(J/ 91IJe)
,-------' \ t-
"
\ ( ~-------
,
\ I \
'\ // \. Blu;t:'·d) R.dRaie
ofFI,.
\ / \ ?- PerMan
~ +1 \ I .'
t c rce rzcno " \ I ;
I rf~JgE}f \
I ". _/' ) t all~t K:1
I / ' l',ng leoto ...+-.,
/ \ I
i \ 1 / \ H- ",
Col/fiction / /
i \ I I \ I .'
. .and;';o:::';"~
Command
/ andContro/ \ I I \ \
/ "\
i \ , ++ I /
I \ bl ;jl ~ I o s: f?0 (j'tl + f?l· cj '?, )s•.H ',1e ~ ~?tO'd Re se rv e
.. • Re.-JPese rve s
----------- \ f~\l t':t . Cornt.ot . . . - - - I\m vCll r~()t d ~ C o rn m l!m 0 rll Rq t,:,: Rom o ,nH"'!g
1-
.c \ ~
\ I
I -,
,
\ I
\ I
/ -,
/
T '~
/ R.g:;,n:~:~)
I +'\ " +-
+ 81 u" '<~"'<v" -+ IJIUA'" ! ?(~ Cl LC5~
H • A rrlV ~jl R(lt..~ - -- "C~rT lb()t
® 7a te
••
\ +1 \
\ I \
\ I \
\ I \
/ \
\ + \
---r-,
" ( R.d,/BlUfI .')
, _' ~.-'r:~:~ ~~,':, -/ Shol
BIwRat. / "'---_.----.
oIF".P., -
Man
" Iniliol
O -_._ - _ ./
ReJerve Fronh9'
': ":)al
FOlee Fore.
~
,t
,
_ Tolal
_, __. Forc. - - ./
-----.,---"
Fig. 7.11 Pressure points in the simple combat model.
......
-----------------------
" ...
""
20 .
/
---, /
/
"' . . . _........"
10 ..-rt---......
TIME (HOUR) SIMlA.ATED TIME

(BLUE) BLUE FORCES IN CONTACT wITH RED
(RED) RED FORCES IN CONTACT WITH BLUE
60 .
so. ---\
\
\
\
\
40 . \
\ ,
\
\
30.
i'7---------------
\
I'\,
20.4+-------4
\
\
,,
\ ,
10 . \
,
\
,
\
\ ,
TIME (HOUR) 5IM~TED TIME
(BLUE) BLUE RESERVES REMAINING
(RED) RED RESERVES REMAIHING
(1) SWITCH TO TRIgGER COMMITMENT OF BLUE RESERVES
A - BASE CASE
Fig.7.12 The combat model's pressure points (output) .
less, case D is an improvement over the base case, even though the behav-
iour mode has gone back to being the same as the base case's.
Since Case C is very favourable to Blue, and Case D is at least to his
comparative advantage, it is intuitively obvious that the two combined will
be even better than either in isolation. Figure 7.12E shows the phenomenon
60 .
50 .
40 .
30 .
TI ME (HOUR) SIMULATED TIME
(BLUE) BLUE FORCES IN CONTACT WITli RED

(RED) RED FORCES IN CONTACT WITli BLUE
60 .
50 .
40.
30.
,------------
I
I
TIME (HOUR) SIH'.LlTED TIME

(RED) RED RESERVES REMAINING
(1) SWITCH TO TRIGGEIl C0144ITMEHT OF 9L.UE RESERVES
B - INCREASED BLUE TRANSPORT CAPACI TY
Fig. 7.12 Continued
of counter-intuitive behaviour which we saw in the consultancy problem in

Chapter 6. Blue is now winning at TIME =60 , but only just , with 3100
survivors whereas Red has 2400. This is worse than case C, though better
than D, and the behaviour mode is different from the two previous ones of
60.
40.
30 .
20.
---', "''''.,,----
10 . ' ..... - ........
TIME (HOUR) SIMULATED TIME

(BLUE) BLUE FORCES IN CONTACT NUll RED
(REO) RED FORCES IN CONTACT NUll BLUE
60 .
50 .
40.
30 .
20 .. ++ -i
10 .
TIME (HOUR) SIIIlLATED TIlE

(REO) REO RESERVES REMAINING
(I) SNITCH TO TlUlillEA CO"'ITIENT OF BLUE RESERVES
C - REDUCED RED TRANSPORT CAPACITY
Fig.7.12 Continued
a decisive Red or Blue victory, respectively. Again, we see the strange

behaviour of the Blue commander in committing some reserves slightly
later at TIME = 14, but having enough left to turn the tide of combat again
at TIME=40.
,,----------------------_.
20 .
---_
/ /
".... ....
,, .-
.- .-
10 ."'IT-~~ ...- - ........

(REDI RED FORCES IN CONTACT WITH BLUE
2'1 60 .T]
1.7sj :1
J
50 . 1
1.~ 40.
I
I
I
1.2aJ
I
1
1. :30. r-----------
0.7~ I
20.++------1. I
o.!ij I
0.2~ 10 . I
I
o.J 0.=-+._...,.....,....
I *101· . TIME (HOUR) SIMULATED TIME
o.
!--RRES
I----FRTRIG
LSLRES (BLUE) BLUE RESl:RVES REMAINING
IRED) RED RESERVES REMAINING
(1) SWITCH TO TRIGGeR CO"'ITMENT OF BLUE RESERVES
D - REDUCED BLUE MOVEMENT DELAY

Fig. 7.12 Continued
Policy as a pressure point

The strange behaviour of the Force Ratio Trigger in Fig. 7.12E suggests that
the Blue commander's 'policy' of committing reserves only when he needs
to do so is influencing the behaviour in unfavourable ways. Let us, there-
---,
'-M"-~~
" ......- - ---'
TIME (HOURI SIMULATED TIME
(SLUE) SLUE FORCES IN CONTACT OITH RED

(RED) RED FORCES IN CONTACT OITH SLUE
60 .
50 .
40 .
30 .
20 .++- -i.
10 .
TIME (HOUR) SIMlA.ATeD TIME
(SLUE) SLUE RESERVES REMAINING

(RED) REO RESERVES REMAINING
(1) SOITCH TO TRIGGER COMMITMENT OF SLue RESERVES
E - CASES C AND D COMBINED

Fig. 7.12 Continued
fore, alter the model so that, once he has started to commit reserves, he
commits them all. To do so we need to transform FIG7-12.COS into FIG7-
13.COS.
The change lies in one sector , shown below, with explanations following
the equations.
note
note amendments to Blue commander's decision on
note reserve commitments
oo~ =========================
note
note the original equation
note
note a frtrig.k =clip(l ,O,frat.k,bcafr)*wswitch.k
As stated, this is the original equation, for ease of comparison . It correctly
does what the Blue commander wanted to do, that is to commit reserves
when FRAT.K equals or exceeds BCAFR and the war has started. Its side
effect is that, if FRAT.K goes back below BCAFR, then reserve commit-
ment will stop, but will start again if FRAT.K goes back above BCAFR.
This pattern will be repeated whenever FRAT.K changes relative to
BCAFR.
What we require is to keep FRTRIG.K at the value of 1 once it has
changed from 0. Since this book aims to be as independent as possible of
any particular software package we shall solve the problem using only the
common system dynamics formats .
note
note revised equation using only stand ard System
note Dynamics equations
note
I frtrig.k =frtrig.j +dt*rcpul.jk
n frtrig=O
The key is to realize that FRTRIG must 'remember' that it has been
activated on the first occasion that reserves are committed. This 'memory'
°
requirement means that we must use a level equation in which the dummy
rate , RCPUL, will act as a PULSE to move FRTRIG.K from to 1 within
1 DT.
r rcpul.kl =clip(l/dt,O,frat.k,bcafr)
x *clip(l,O,O,frtrig.k)*wswitch.k
The first CLIP makes RCPUL equal to l/DT as soon as FRAT.K equals or
exceeds BCAFR. Of course , if FRAT.K stayed above BCAFR, then
RCPUL would be l/DT all the time, so FRTRIG would move from to 1
to 2 to 3, etc. with each DT for which FRAT.K equalled or exceeded
°
BCAFR. What effect would that have elsewhere in the model?
To prevent that effect, and to keep FRTRIG to its logically correct value,
the second CLIP permits only one PULSE, since it only produces the value
1 when 0, the third argument, is greater than or equal to FRTRIG. In other
words, it will only allow a change to FRTRIG when there has been no
previous change. The use of WSWITCH prevents any change until the war
has started.
Summary 231
d rcpul=(I/week) pulse to turn on blue commander's switch for
x commitment of reserves
As always, new variables must be defined .
°
The reader should run a small model , essentially consisting of only these
few equations, and with FRAT.K made to STEP from to 3, say, and study
what happens to FRTRIG with the two different policies.
note
note revised equation using COSMIC FLIP function
note
note a frtrig .k=flip(frat.k,O,bcafr,O)
Readers who have access to COSMIC can achieve the same effect more
simply by using the FLIP function, described in the COSMIC User Manual.
When this change is implemented in the model , and the model is run to
reproduce those cases in Fig. 7.12 in which FRTRIG was turned on and
then off by the old policy, the results shown in Fig. 7.13 are produced. In
case B, the overall outcome is not changed, the only difference is that
FRTRIG, once turned on, stays on. In case D, Blue lasts about 6 hours
longer and Red has 16700 survivors, rather than 22800. Case E now ends in
a complete Blue victory, with 17300 troops left. Notice , in the upper half of
case E, Red's reserves are still obligingly marching up and being promptly
destroyed, from TIME=31 until the last reserves arrive at about TIME=45.
This demonstrates the working of the non-negativity constraint on Red
losses.
Even in the original model, Blue 's policy can have strange effects. As
with the Raw Materials Manager in the DMC problem, the Blue com-
mander takes no account of the reserves moving forward , or of the Red
reserves, whom we shall assume he can see in the distance. Define two new
variables, BRER and RRER, for Blue and Red reserves en route, respec-
tively, write suitable level equations, and alter the equation for FRAT.K to:
a frat.k= (red .k+rrer.k)/(blue.k + brer.k+ beps)
and see what the effect is.
SUMMARY
This chapter has shown the immense range of design possibilities which are
usually to be found in system dynamics models. We examined two models,
one in which the behaviour arose from the system's response to an external
driving force and the other in which it derived from the interactions within
a completely closed system. In the first case, we experimented by changing
parameters and structures, in the second , only parameters were changed.
We used the term 'pressure points ' in that model to suggest that parameter
TIME (HOUR) SI~TED TIME

(RED) RED FORCES IN CONTACT WITH BLUE
60 .
50. ---,,
,,
,,
40. ,,
,,
,
30 .
t\---------------
I '\,
,
I '\
I '\
I '\,
,
TIME (HOUR) SIMU.ATED TIME
(BLUE) BLUE RESERVES _ININIl

(1) SwITCH TO TRI-.. ClMlITIEHT OF BLUE RESERVES
A - INCREASED BLUE TRANSPORT CAPACITY

Fig. 7.13 Combat model - once committed.
Summary 233
60 .
50 .
40 .
30 .
1
I
1
20 . 1
1 //-----------',-- _
I /
I , ~/
10 .-t1-t-- ............ __ ........
I
1
I
o .~I-k---\,---,.,._-ln.-tt-I,rn-~6..._~",.,...~=____itrr__\1ro:_,1'r;..____I::
*10'1 . o.
iL
TIME (HOUR) SIMlkATED TINE
BLUE (BLUE) BLUE FORCES IN CONTACT WIlli RED
- -RED (REDI RED FORCES IN CONTACT WIlli BLUE
60 .
---,,
,,
,,
,,
,,
,
f\---------------
I '\
20 . # - - - - - \ \
I '\
I '\
I '\,
,
TIME (HOURI SIMl.LATED TINE

(1) SWITCH TO TRIGaEA CO"IITNENT OF BLUE RESERVES
B - REDUCED BLUE MOVEMENT DELAY

Fig. 7.13 Continued
234 Policy experim ents with system dynamics models
..... ""
""
""i'_--L-~=-u.v
30. 35 . ao.
J
'15 . So. 55 . so.
TIME (HOUR) SINlA.ATEO TINE
(BLUE) BLUE FORCES IN CONTACT WITH REO

(REO) REO FORCES IN CONTACT WITH BLUE
60 .
50 . --- ,
"" ,
"
40 . ""
" ""
" ""
"
I- ~- - - - - - - - - - - --
.-H-----~
I ''' ' ',"
""
"-,
""
""
""
""
TIME (HOUR) SINlA.ATEO TINE

(REO) REO RESERVES REMAINING
(1) SWITCH TO TRIGGER COHHITNENT OF BLUE RESERVES
C - CASES C AND D COMBI NED

Fig. 7.13 Continued
Summary 235
changes might require investment in different aspects of the system. To

emphasize the difference, a parameter such as WAP in the DMC model
can, in the real world, be changed simply at the behest of the managers, if
they deem it to be in their interest to do so. On the other hand , a factor such
as BMDEL, which is a parameter in the model , can only be changed, in the
real world, by investing limited resources to do so. It seems useful to
distinguish these costly parameters by a special term.
Naturally, the distinction between driven and undriven models does not
mean that only the latter have pressure points and only the former have free
parameters. In the combat model, FRTRIG is a free parameter and, in
DMC, to change DDEL would probably require considerable time and
expense to set up a new supplier network.
The whole point of this chapter is, we reiterate, to stimulate thinking
about the wide range of changes which are possible in models and the
dramatic effect that changes to parameters and, more especially, structures
can have on system performance. The reader should now experiment fur-
ther with these models, others developed in the previous chapter and the
new problems in Appendix B.
However, it should also be clear that there is a limit to the amount of
useful information about how to improve a system's performance than can
be gained by simply experimenting with a few parameters at a time. Experi-
mentation is essential to deepen insight into what can be done, but we
require a more powerful method of searching the parameter and structural
possibilities. That power is provided by dynamic optimization, and it is to
that topic that we turn next.
CHAPTER 8
Optimization in system
dynamics
INTRODUCTION
In the last two chapters, we built and experimented with several system
dynamics models. The problems in Appendix B call for more model con-
struction and exploration of model behaviour. The conclusion is that the
behaviour of a model can be changed quite remarkably, for better or worse,
by changes to parameters or pressure points. Usually , even more significant
variations in behaviour stem from changes to the model 's structure. The
drawback is, however, that there is always a nagging doubt that, had one
tried only one more experiment, something even better would have been
found. Unfortunately, there is always yet one more experiment, so the
process never stops .
It would, therefore, be highly desirable to have some automated way of
performing parameter variations,' up to a certain number, and reporting to
the analyst the best result found in that set of experiments. On the face of
it, that could be provided quite easily by writing software which would test
combinations of parameters and report the best result. In fact, that is not
what is needed, because the number of possible combinations and conceiv-
able numerical values of the parameters is usually colossal; in theory, it is
infinite. One needs, therefore, some sort of guided search of the parameters
to be considered and the numerical value each might have so as to seek out
the result which is most rewarding in terms of enhancing the system's
performance and not pursue blind alleys. Unfortunately, there is no perfect
way of achieving that, but the principle of dynamic optimization comes very
close to providing this subtle searching of the design possibilities of the
system. This chapter will discuss the theory of dynamic optimization, illus-
trated by a simple example; the next will apply it to some of the models
already developed.
It will be noticed that the first paragraph referred to 'experimentation'
and the second has mentioned 'design'. This is the distinction which was
1 Throughout this chapter, the term 'parameter' includes the structural parameters which were
discussed in Chapter 7.
The system response and the parameter plane 237
mentioned in Chapter 7 between finding the range of behaviour of which a

system is capable, and deepening one 's insight into it, by experiment, and
designing policies and structures to give the best performance which can be
attained. An important aspect of dynamic optimization is the development
of measures of system performance: so-called objective functions . We shall
have more to say about that later in the chapter.
Unfortunately, most of the system dynamics software products summa-
rized in Appendix A do not support optimization. The two which do differ
very considerably in the details of how they work , though they adhere to the
general principles to be described here. In this book, we shall use the
COSMOS optimization software to illustrate the underlying principles and
the practical power of optimization.
THE SYSTEM RESPONSE AND THE PARAMETER PLANE
Figure 8.1 shows the first of the ideas we shall need to understand. The
diagram shows a two-dimensional picture of a three-dimensional object.
The two dimensions in the horizontal plane are labelled for two of the
parameters in a model. It is essential to realize that these may be ordinary
policy parameters, such as WAP in the washing machine model; they may
be structural parameters, such as ALPHA and its cousins in the same
model ; or they may be pressure points , such as BMDEL in the combat
model. Each parameter has a range within which it may lie, shown as, for
example, P1uPPER and P1LOwER. The initial, or base case, values are labelled
P1! and P21 on the parameter axes. When these two values are projected
into the 'parameter plane', following the dotted lines, they meet at Point A.
When the model is run with those values, the response is, we shall suppose,
rather poor, so a short line is drawn in the vertical direction to indicate that
poor response. Naturally, this idea of the model 's response at a point in the
parameter plane is valid whether there are 2 parameters, 22, or any other
number.' However, with 22 parameters one would need to draw a 23-
dimensional object on a two-dimensional sheet of paper, which is rather
difficult.
The third, vertical, axis is a measure of the quality of dynamic behaviour
which the model produces for any given combination of parameters. For the
moment, we shall simply label that scale as running from 'Bad' to 'Good'.
Later in this chapter we shall explain how 'bad' and 'good' can be
measured.
If the two parameters are now changed to P12 and P22 , a new point B is
defined which, we shall imagine, produces better performance and hence
qualifies for a rather longer arrow in the vertical direction.
2 If there are more than two parameters, the 'plane ' becomes a hyper-plane in n-dimensional
space. We shall stick to the simpler term of 'parameter plane '.
"0
Q
Q
e
~">$'~"
~",__ _____ !~~~rJat

----
, PointB
..
"0
..
<>,,~o"~
- - - - - -,~-.>
Plane ,
__
--_
t_ Response at
Point A
=
~
The parameter " , , :: Point A
, ,
p P2,
liraIneter 2
Fig. 8.1 The parameter plane and the model's responses.
,I
\
Edge of Bose of
Response Surface
in the Parameter
Plane
Fig. 8.2 The response surface.

Hill-climbing optimization 239
Figure 8.2 now extends the argum ent to suggest a response surface in the
three-dimensional space. The response surface is the surface traced out by
the responses at all possible combination of PI and n. The heads of the
arrow s at Points A and B just touch this surface and can be imagined to lie
on local contours of the 'hill '. The diagram tries to show that the response
surface is a very rugged mountain, with several peaks. Two of the peaks , a
and ~ , are of rather different heights and a is clearly 'bette r' than ~ . Figure
8.2 also suggests that there are all sorts of irregularities and gullies and that
the mountain slope is much steeper in some places than in others. It helps
to understand this (and Fig. 8.4, which we shall encounter later), if one
looks at a photograph of a mountain range. The reason for the extreme
irregularities lies in the complexity of system dynamics models and the non-
linearities which they often contain.
We have just argued that we do not wish to test all possible combinations
of parameters but , unless we do, the shape of the response surface will be
unknown. The problem of dynamic optimization is: how does one find one's
way to the top of a mountain of unknown shape?
HILL-CLIMBING OPTIMIZATION
The answer lies in the analogy of the blind man who is marooned on a
mountain and wishes to find his way to the top.' His strategy is to feel the
shape of the ground around the point where he is sitting (Point A) . Having
detected the direction in which the ground slopes up most steeply in that
vicinity, he takes a cautious step or two in that direction and then feels the
ground again. In this way, he hopes to find the top of the mountain, as
shown by the sequence of arrows moving up the surface of the hill from
Point A in Fig. 8.3. Unfortunately, the blind man 's strength eventually fails
him and he can go no further than 11. This is not even as high as ~ , so there
is no guarantee that the blind man will reach his goal, a , or even another,
lower, peak.
The idea of feeling one's way and taking steps in search of the top of the
hill corresponds in optimization software to iterations of the model. Each
iteration is a run of the model with a particular set of values of the collection
of parameters being searched. The first iteration uses the base case para-
meter values. After the model has run, the value of the objective function is
calculated as a measure of how good or bad the performance was. The trick
comes in 'remembering' which sets of parameter values gave good results
and using that information to predict how the set of values should be
changed to carry out the guided search for good values, and not wasting
effort examining parameter values which lead nowhere.
J A blind woman would be far more sensible and would try to find her way to the bottom, and
safety. She would, however, use the same principle.
240 Optimization in system dynamics
a Th
Path/of
e Steepest Ascent
I
,' "
1\ I "
... ' I
I
/\ , I
: J.l" ...' \ The Response Surfoce
'I \
\,I l'
L '\ . /
I "" T ,
'1/
I " ,
I ~ "
,' ' / '- Local Contours
I <>\"",~.. , / - . , - - for Responses
;
...e.'II\ ~\ ,
~ I 'xatAandB
',
~e. , -___ ~ \
.: ,'------- -:':1'" \
~ ~\, ,' - - - - \- Point B '-
\v><¥'l-' ...... "',""': ... "

~ r plane, ---- ,
. ,' Theparamete ,,' -: Point A '.
... ... ' , ".
I
\
Edge of Base of
Response Surface
In the Parameter
Plane
Fig. 8.3 The concept of hill-climbing optimisation.
The trick derives from the use of hill-climbing algorithms developed in

the mathematics of numerical analysis and adapted for use in system dy-
namics software packages. A number of good, efficient, algorithms are
known. The reader interested in the mathematical technicalities should
refer to the literature on numerical analysis; the pragmatist can rely on the
evidence that the approach works, which will be shown for a simple exam-
ple later in this chapter and developed more fully in the next.
The equivalent of the blind man's strength failing him before he reaches
the top of the hill is not commanding sufficient iterations to be performed.
Since, however, the COSMOS software is very efficient, there is no harm
in demanding 500 iterations, if one wishes. Experience indicates that, for
many models, about 30 iterations are sufficient to find dramatic improve-
ments in performance. Even with a large model, a hundred or more itera-
tions take no more than a couple of minutes on a reasonably powerful Pc.
lt is important to understand that hill-climbing by the method of steepest
ascent is a heuristic method; that is to say, a method which is based on a
common-sense rule of thumb which is not guaranteed to produce the
Overcoming the limitations of heuristic algorithms 241
optimal result. This is in distinction to the well-behaved mathematical

functions, for which a guaranteed optimal result can be proved to exist and
can be found by such standard methods as calculus or linear programming.
OVERCOMING THE LIMITATIONS OF HEURISTIC ALGORITHMS
Computer software is not limited by the blind man's strength, so that any
number of iterations can be commanded. Even so, there is no guarantee of
finding the maximum of the response surface. However, numerical hill-
climbing algorithms are sophisticated and are capable of searching with
something close to the intelligence that the blind man would use.
Consider Fig. 8.4. The vertical axis is still the model's response, but the
horizontal axis should be imagined to be a 'cross-section' of the parameter
plane, representing varying combinations of the two parameters. The hill-
climbing search has started at Point A and followed a steep ridge until it
reached B. From there, a valley leads forward into the hills, and there is a
ridge to the left, but the locally steepest direction is towards C. On reaching
C, it becomes clear that it is a false peak, from which all directions lead
downwards. Since the man is not yet exhausted, he would remember C as
the best point so far discovered and, reasoning that he can always go back
to C, he accepts the loss of height and searches towards D. At that point, he
discovers a slowly rising path which emerges from behind peak C, as shown
by the faint line, and moves off to E and hence, we hope, to F. As soon as
any point is reached which is better than C, it will be remembered as the
point to which he can return if no better solution is found.
/
1/7'\
I
F
"
/
A new dIrection
/a pea
A t lee k
. ' I \, E C
/ J ~-_ /--_
/ I' I \ ..... - -:""7"'."._"'__'... . .-.D_
/ / \ I
I I , I
I / ~B
I / II
1 I 1 "-
I
/
I X
A I
I "A local ridge an the mountain
An oltemattve "-
'!rating paint ~
.
'l:l
ell
The starling pain!
A 'Cross-section' ofthe Parameter Plane
Fig. 8.4 Continuing to search.

There is still no guarantee that this course of events will unfold, so

another strategy for avoiding failing to find the real peak would be to
recommence the search at a new starting point, X, in the hope that there is
a ridge going directly to F. In practice, the improvement in behaviour found
through optimization is usually so large that it is hardly worth the effort of
repeating searches from new starting points. The main benefit of doing so
would be to increase the experimenter's confidence that a good solution has
been found.
THE PERFORMANCE OF A SIMPLE MODEL
Problem 9 in Appendix B is intended to give practice in developing a

model , but it will also serve as an example of optimization. Figure 8.5
(FIG8-5.COS on the disk) shows the base case and is practically the same as
Fig. S.9 (we hope that you are doing the problems as you work through the
book!). The addition is that Desired Inventory, DINV, has also been plot-
ted and the plotting scales have been changed, as we no longer need to show
negative behaviour. Although the model now avoids the error due to the
omission of the non-negativity constraint, its behaviour is clearly fairly
disastrous. DINV and INV do not match until TIME=140, which is 2 years
after the shocks in CONSR. Furthermore, PMBLOG rises to a peak of
about 1100 at TIME=30, but then falls to zero at TIME=55 and stays there
for 45 weeks. This corresponds to the factory having to be closed for nearly
a year. All in all, a less than sparkling performance, but by no means
untypical of the behaviour of real systems.
200
15oo-m "
1250~ 175
;' " \
I.,I :I 150
:
"
\
1000 I.!:: .,1 \

ill
!I! ./ ~I
\
",
125
750~ 100 II~

II: 75 R ~\ '\
500~ \ 1\\ "
'1 50
++i -, ~\ \\
1
1
\ \ '
250
25
I.'1: I
-,
-
r-:~~-
oJJJ 0 . . ~o . 30 . ~o ""go- . Bo . 70 . Bo . !to. 1100!110 !'20!'30 !'40.1'50.~60 .

III
·,: L TI ME (WEEK) SIMIJl.ATEO TtHE
I ---------bNV CONSR (PARTS/MEEK) COHSUIIPTIllN RATE OF PARTS

(PARTSI ACTUAl. INVENTORY OF PARTS
C-=--==,pA~~OG l~:~~l gm~ f:~~~YF: ::iiACTlJllE
BASE CASE (FIGURE 5 .9)
Fig. 8.5 Simple model from problem 9.

Formulating an objective function 243
How, though, can we measure that poor performance so as to have an

indicator of improvement?
FORMULATING AN OBJECTIVE FUNCTION
As we remarked earlier, an objective function is a measure of system

performance which is used to guide the optimization search. It should take
account of what the system is trying to achieve and calculate the extent of
its success. It has to be admitted that developing objective functions for
system dynamics models is something of a black art, and that this is an area
which is ripe for much more research.
Simple approaches based on common sense do, however, work well and,
in this case, it seems reasonable to assume that the management objective
is to make INV match closely to DINV, otherwise there seems little point
in having a desired inventory. If, however, we wrote:
A OBJFUN.K=DINV.K-INV.K
we would mislead ourselves. The COSMOS package works by taking ac-
count of the value of the objective function only at the end of the run, when
TIME=LENGTH. (Other packages may operate differently, though the
underlying theme is the same.) It would be no good to optimize only for that
one value of TIME, and we must take account of the behaviour over the
whole of the run. This requires a level variable which will act as a memory
of what happened during the run , and we might define an Inventory Pen-
alty, INVPEN, as:
L INVPEN.K= INVPEN.J + DT*(DINV.J- INV.J)
N INVPEN=O
This will not do, however, as there is clear evidence from Fig. 8.5 that the
difference between DINV.K and INV.K is sometimes positive and some-
times negative. To avoid the positive and negative discrepancies simply
cancelling each other we square them and use:
L INVPEN.K=INVPENJ + DT*(DINVJ - INVJ)* *2
N INVPEN=O
INVPEN is, of course, the area which would lie under a graph of
(DINV- INV)2.
The numerical value of INVPEN will be some strange number, so, to
make performance comparisons easier, we introduce a new variable ,
OBJFUN.K, defined by
A OBJFUN.K= INVPEN.K/SCALE
where SCALE is chosen to make OBJFUN, the objective function we shall
actually use, equal to 100 at the end of the base case run . In this case,
244 Optimization in system dynamic s
SCALE is equal to 46.7933 X 104 which is found by running the model with
SCALE=l and the base case parameters, observing the value ofINVPEN
at the end of the run, and calculating SCALE accordingly. In this instance,
it is evident that we wish to minimize OBJFUN, which is tantamount to the
blind man finding his way down the crater of a volcano (hopefully extinct)
by the method of steepest descent.
These changes are added to the model to give FIG8-5M.COS on the disk.
THE SIGNIFICANCE OF THE OBJECTIVE FUNCTION
It is very important to be quite clear about the significance of objective

functions.
In the first place, they are extra equations added to the model for the
analyst's benefit. They are not part of the real system and have no physical
meaning. They are only there to help the analyst, and the software
which serves him, to keep track of how improvements to behaviour can be
found .
Secondly, the dimensions of the objective function do not have to have a
real-world meaning. The dimensions of INVPEN are [WEEK *PAR'f2],
which does not correspond to anything in the real system. For this reason ,
the dimensions of INVPEN and OBJFUN in FIG8-5M.COS are given as
[ARBUNIT] to denote arbitrary units with no physical meaning, though
one could equally well use [WEEK *PART**2]. If ARBUNIT is used,
dimensional analysis, whether by hand or by software, will suggest that the
equation is dimensionally invalid, but that can be ignored.
Obviously, there is nothing wrong in having an objective function which
is dimensionally valid, such as minimizing average inventory over the
course of the run , but the objective function and its components do not have
to obey the strict requirement for dimensional consistency which applies to
all of the rest of the model.
Thirdly, one has to be very careful about choosing objective functions.
On the face of it, minimizing average inventory is an attractive thought, but
a moment 's consideration shows that the true minimum inventory is zero ,
and a firm which has zero inventory is usually not going to sell very much
and could rapidly head for bankruptcy. In other words, choosing an objec-
tive function is not a trivial matter. Simply getting smooth dynamic behav-
iour is not always ideal if the smoothness has nothing to do with what the
firm is really trying to achieve.
OPTIMIZATION EXPERIMENTS
When the optimization software is activated, one has to specify the par-
ameters to be searched and to state the upper and lower values of each one.
In this problem, there are four parameters, TAC, TCI, TICAS and
Optimization experiments 245
TEBLOG. TTCAS is a gain, the rest are delays. To start with, we allow all
four to be in the parameter plane to be searched. All the base case values
are 6, and, after careful discussion with the management of the company, to
see what kind of changes they might be prepared to tolerate, we allow them
all to lie in the range from 2 to 10. The software is instructed to do 30
iterations. Beyond that , the precise details of how COSMOS works need
not concern us; the results are what we are interested in.
The optimal values of the parameters are reported to be TAC= 10,
TCI=2, TTCAS=2 and TEBLOG=2. The value ofOBJFUN falls from 100
to 4.03. This reduction of practically 96% is by no means unusual. All the
parameters have been driven to their extreme points.
It is interesting to note that the gain, TTCAS, has been reduced, but that
only one of the delays, TAC, has been increased , while the other two delays
have been reduced. In Chapter 7 we stated that it is a rule of thumb in
control engineering that reducing gains and increasing delays is likely to
increase stability. It is indeed true that that is the rule of thumb, but it has
clearly not worked in this case. Blind adherence to such rules can be
seriously misleading.
It is usually not a good idea simply to throw all the parameters into the
optimizer and take the results on trust. To do so is to abandon thought
and rely on computation. In this case, we might feel that to increase TAC
too much might make the system too insensitive to the unpredictable
changes in CONSR, so let us try another optimization in which only TCI ,
TTCAS and TEBLOG are allowed to enter the parameter plane, with
the same ranges as before. The results are that all three are driven to their
lower value of 2 and OBJFUN is reduced to 5.268. On the face of it,
this is nothing like as good as the previous case, because OBJFUN is
about 30% larger. In fact, the visible differences are very slight, as shown
in Fig. 8.6 (FIG8-6.COS). The only difference between the two plots is
that INV reaches a maximum of 562 with four parameters and 581 with
three.
The optimization with three parameters has driven the value of TTCAS
to 2, which means that the firm is trying to operate with only 2 weeks of
stock cover. This is fine when demand is constant, but might be quite
insufficient for any noise in the demand pattern and provides little protec-
tion against another upsurge in CONSR. When it sees these results, man-
agement therefore feels that it has been a little rash in allowing TTCAS to
have a range from 2 to 10, and wishes the optimization to be done with TCI
and TEBLOG ranging from 2 to 10, but TTCAS in the range from 4 to 10.
The optimization produces all three parameters at the lower limit of their
ranges and OBJFUN reduced to 12.84. In management terms, the param-
eter values correspond to holding a prudent level of stock, but reacting
quickly to changes. The behaviour is shown in Fig. 8.6C, and is clearly not
as good as either cases A or B in Fig. 8.6. For instance, the maximum value
of INV is now 857.
246 Optimi zation in system dynamics
1500 Ill 200
1250~ 175
Ii: 150
1000!!:
"fil
,I:
l 125
750~ 100
!Ii
r-.
75
500~
ii I:I: 50 f \
\
250 -m 'II 25_ ''f',;:--_\~\,:---------------
Om 0 '- " , 60 ,
II' L"
' 1i
TIME (lIEEK) SIMU"ATED TIME
CONSR (PARTS/lIEEK) CllN!lUMPTIllN RATE OF PARTS

1 --------- INV (PARTS) ACTUAL INVENTDAY OF PARTS
L-=-==,a~~~OG l::~:l :~~~ t:~~~YF~ =ACTUI£
A - OPTIMISATION OF 4 PARAMETERS
200
175
150
125
100
75 ~\
I \
50 ,
I
I
\
\
\
\
\
\
Fig. 8.6 Optimization of the simple model.
At this stage, and for this very simple model , optimization has now run its
course and management must choose, on criteria which it would probably
find hard to articulate, which of the results it prefers. If, of course, manag e-
ment can articulate its objectives more precisely, a fresh , and more subtle,
objecti ve function could be developed and the process repeated. We have,
however , covered enough of the theory and a demonstration of the practice
to serve the purposes of this chapter; more detail will be seen in the next.
For the moment, we must turn to some final observations on the concept of
opt imization .
Summary 247
1500 Ill 200
1250~ 175
l000ff,'I:
150
125
!I: 100
I
~\
\
750+H I \
!I:
I
I
\
,
,,
75 ,,
500tJ-j
,,
I'll,
,
50 ,,
,,
250+H
'I'
25
om
,'
0 L \
--~----------
ao , L50, I.:r-L-'L'L'h'L'
L ! ' k'T'T"
so , 10 , eo, 90 , 100 ,,10 ,120
II' L
, , 2 0 , ~O ,
TIME (MEEK) SIMULATED TIME
,130 ,,40 ,,50 ,' 60 ,
' 1: CONSR (PARTS/MEEK) CONSUMPTION RATE OF PARTS

1 -- ------- INV (PARTS) ACTUAL INVEMTORY Of' PARTS
C-=--=='~A~~OG l::~~~l ~~~~ ~~~~~YFg:; :~ACT_

C - HIGHER STOCKS
Fig. 8.6 Continued
SIMULAnON BY REPEATED OPTIMIZAnON
The concept that the top of the hill is sought by repeat edly running
the model makes it clear that the technique is optimization by repeated
simulation. However, the two oth er cases, with reduced numb ers of pa-
rameters and with different ranges should make it clear that the real und er-
lying theme is sim ulatio n by rep eated op timization . The model is opt imized
and something is learned. That leads to furth er optimization and more
learning , and so on. The power of the optimization calculation is to provide
a much more powerful guided search of the parameter space than could
possibly be achieved by ordin ary experimentation , but the principal aim is
still to experiment and to understand so that a better experiment can be
designed .
The idea of simple experimentation to see what the model might be
capable of is a fundam ent al precursor to optimization. One cannot devise
an intelligent objective function without first having 'played' with the
model , and an unint elligent objective function is worse than useless.
SUMMARY
Th is chapter has covered the underlying theory of optimization as applied

to system dynamics models and showed something of the power of the
approach by a simple example. The main point to grasp is that optimization,
for all its superficial attractiveness, is not a panacea and it does not guaran-
tee good analysis. There are, indeed, some limitations to the approach.
The first limitation is that the hill-climbing algorithm does not guarantee
an optimal solution. In practice, that matters less than might be thought
because most managed systems perform so badly that any improvement is
welcome and the differences between different optima are usually much
less than the objective function values might imply, as we saw in Figs. 8.6A
and B.
Secondly, the optimization technique does not, of itself, give any guidance
on the development of a good, subtle , objective function. A poor objective
function, such as minimizing inventory, might be truly disastrous.
The final weakness is that the thought of optimizing something is so
seductive that the narve analyst might stop thinking .
In practice, it is only the second and third limitations which are serious
and, provided one thinks , optimization is probably the most powerful devel-
opment in system dynamics since the field was first developed.
CHAPTER 9
Optimization in practice
INTRODUCTION
In the previous chapter we examined the theory of optimizing system

dynamics models, drew attention to some of the limitations of heuristic
algorithms and suggested how they might be overcome, and illustrated the
argument by optimizing the behaviour of a very simple model. We placed
great emphasis on the idea that optimization is not a panacea which ab-
solves the analyst from the need for thought about the problem, and we
suggested that the correct point of view is to think in terms of simulation by
repeated optimization rather than optimization by repeated simulation. In
this chapter we shall take that line of thought a long way forward by
studying, in detail, the optimization of two of the models developed in
earlier chapters .
In particular, we shall examine the problems of unconstrained and
constrained optimization. The difference between these two forms is that
in, for example, the DMC model, parameters representing management
policies and the structure of the information flows can be changed without
limit or cost if management chooses to do so. The question is, therefore, to
identify the changes which will give the greatest benefit to the performance
of the system. In the combat model, on the other hand, money must be
spent to bring about changes in the various pressure points we detected.
The question is still to identify the changes which will give the greatest
benefit to the performance of the system, but without exceeding given
financial resources.
THE IMPORTANCE OF EXPERIMENTATION

BEFORE OPTIMIZATION
We studied the DMC problem in some depth in Chapters 6 and 7. In

particular, in Chapter 7 we explored the wide range of behaviour which that
model can produce and we saw that there is a large number of possible
changes one can make to this system. Some of the changes are simple
parameter variations to test the effects of alternative forms of a given
policy; for example, increasing TARMS, the time to eliminate raw material
stock discrepancies, corresponds to management taking a more relaxed
250 Optimization in practice
attitude about mismatches between desired and actual Raw Material

Stocks. Other changes, such as making ALPHA equal to 0, correspond to
switching management attention from one information stream to another.
Yet other modifications, such as setting SWITCH1 to 1, create new sectors
of the model, in that case reflecting the purchase of the management infor-
mation system. We also saw that , if, say, SWITCH1 is set to 1, it might still
be advantageous to change the values of ordinary parameters so as to get
the best out of the MIS.
In short , this model, like most system dynamics models has an enormously
rich seam of options to investigate and one would not have detected that
range of options without first having 'played ' with the model, as we did in
Chapter 7. Playing is an important part of any learning, and that is as true
in system dynamics as it is in childhood . Thus, optimization should not be
used until one has fully understood the range of options which exist in the
parameter plane. To do otherwise is to hope that the numerical power of
the optimizer can substitute for the intellectual power of insight and imagi-
nation ; that is a false expectation which will lead to far poorer results than
can be had by combining thought with calculation.
To emphasize this point, Figs. 9.1 and 9.2 reproduce Figs. 7.1 and 7.8
showing, respectively, the original influence diagram for DMC and most of
the options seen in Chapter 7, though we suggested in Fig. 7.10 that there
may be yet more options to be imagined. Even so, the difference between
the two diagrams is quite striking and the reader should review them before
proceeding. The heavier lines used in Chapter 7 to indicate structural
alternatives have been deleted from Fig. 9.2. One loop is shown in heavier
lines for reasons which are explained later.
We do not know, however, whether the improvements we have found,
such as in Fig. 7.9, are the best that can be achieved. It is the purpose of
optimization to help us to be sure.
THE DMC PROBLEM
Having thought about the options which exist in the parameter plane , it is
time to turn our attention to constructing an objective function for DMC.
Let us start by using FIG9-3.COS on the disk. This contains all the options
we have already studied but, for the moment, they are turned off in order to
reproduce the base case behaviour. You may find it useful to review the
discussion of the DMC model in Chapter 7 to remind yourself of the effects
on the system of the various parameters and switches.
Criteria of performance
Figure 9.3 shows the base case performance, which we already
know to be rather lamentable. How, though, can we assess whether
A=:=~=d
1ot0ll Avemge
(WEEK)
--. ~~: w..t. at A ~.f~
Produclionin
:1NM~7Ww()
,/'" I I / / J:::SI~s
/" WAP
/ I \ Raee~~jol £ ,/ (WEEll
/ I \ Stocks
I I I \ "" _gPenod
A..._ _ lid. _ -+ DRMS <,
+ (WM SJ
\
I I \ ~ ",,,,/'
\
I I \ \ '" \
f , \ \ ,/ ,/
-- - --- \
I I \ \ ,/ /' -\"--
I I \ \ / /'
\ / '" \ -- .... <,
I I \ I '"
+y
\\ ,/
J . II Av..o ge ,/
I " -,
J D
... ed \ ProdJctlon \ -,
~~ \ ~~ I \
J DBl OG -- .... \ (WMS/W EEI() I \
I (WMS) -, \ A+ I \
I tI '\ \ , \
I / \ \ I ++ \
I I I -I \ I Row Mot6OOl \
_ ~ StOCk ~~Fncy - - - ... \
I I BoJ.og \ I /'
_ k . 01A ...._ Doscrepcncy \ I / (W MS) " '\ +\
I Otd«s . . 0 . _ + Y' J,~) -, \ I
I ~~ 1'/ \ \ ' I + \~Matetlal
Ord er Rate
I (MEIO I \ \ ' ROof Mateo'"
I \ \ , _St~~ks / Y' (w~~~EI()
I I \ \ I ~(wMS) rllJ»toConfICI
IlawMaletkll .....-
Dltc:er
-, . . ~:-PrJud~~ ~ (MEIO
+
\I (\ RWt
\ \ ~MS/WEEI()
I \ ... ROof Motend +
/ An"'ol Rate
\ Boddog at - / RMAR ....f-- -
\ + ~~ / (WMS/WEEI()
\ ~ (WMS) /
/
nme to COffeet
New Ofder Ord« 'ocklog
Jnflo wRate D/KHG"Y
NO
(WMS/WEEI() (WEEKI
Fig.9.1 The original influence diagram for the DMC problem.

- ------ - -- -- -- - - DtNIIetIWlNu
..... ....7JI'
~- " _ _ ~!!.!-
«~~
",, / . . . 8/ < ; / ~
/' " ...... I (\It&l')
-: -- -- - - - - - - - -- -- - --- - ---- - -~-_ ...... ...... I
/' --- ------ -_ ... /
-: --- » > - - ...... -- - ' , /
TAor _
A===- . r: >: // / < ; ... .... . . . ' ... II
/WHO - __ AC:~ - _ - - - - - - - - - _ __ ,/ ... .... ... " ~
.. Role - - - - - _/ MlM«:IoIA~""'" De.. edR(7oAI
~ ~~-
..... ~'" -, <, Mate' ial o n
/ ' ..... (WMSIWl:EK) I ............ ...... ... ...., ::,:::: : ; . . . . . . .
1- --~ ~
'" I \ I / '4 o-ed .... - ..... , (\'MS) , ,
I I I \ / I I R01=:nal (MIlK) "
I \ I 'V ..- ~OOMS ..... \ \
/ I \ ... _ _ I I ./ + (WMS) \ \ \
I I \
A__
.... I \ / \
/ I \ ~ I \,/ \ a--- - ---\ I
I I \ "1,/ .... _ - - - - - - - __ \ . . . : ' I
I I \ \ I / » < _ _ ~ \ I
I I \ \ I " \ ' " \ I
I I \ \ / ",'" I .... .... ' I
I +" , "A_ago '" I , I I
~ed \ ~~~ " I I
Bock"" \ " ~ I \ I
II o l Orders ...... \ ,/ CWMS/WEEK) \ I
I + , I I
I
I
J /
/'
= '.. \ '\ \
I \~
"" J} - .../
/. ,Switch2
,
++
_ . Raw tolatsll o l
\,' \
Switd, J
J+
I / -, \ I I ' __ -- ::-_ ~SJock cr:~aponcy - - - . . . \ I Raw ""'o1tmol
I on Orde ,
I / ~V \ \ I I . / -- ", " (WMS) .... -, +\ + I o.c,opa"lCy
I + .. \ \ I + _.4- - -- EPS3
=~".:: ~).... I .. \J.~.noI +_ (W Mi)
I ":.~ /'" \ \ \ ./ I 0,,,",""' z >
\ \ \ ' ' RowMatefIOI yo ~ t-
I IWIIIO I \ \ \ / _ St~" ...- ...- (W"""WEEK) , -, ,
I I \ \ \ , // ''''''''' ,=-~ "...- "'- I
I I \ \ \+ -rA:""
\\ , ,
~ I
,
\ I
, ' .... .. "oduc"""
- ~"It' ~- /lOIIlO \ \ " , ""
_"e.-d
_ I
I \ + ":: \ '= I
\ \ ~(WMS/WEEK) I r...-SO
I
\ \ / ~~~ + I I
, 8ock",, 01 - I ( W =EK) I I
I
I\ /
+ ~m
~(WMS' II ~ II /
I + /
~~ - ~ ~~ / " '"
'"
Inflow Rote
New O'de' I fAat
~~r = -----
(WM3~£EK) CWEEIO
Fig. 9.2 Option s for the DMC problem.

The DMC problem 253
250
225
200
50
25
70 .
TIME (WEEK) SIMULATED TIlE
IWMS/WEEK) NEW ORDER RATE

(wNS/wEEK) PRODUCTION RATE
IWMS/wEEK) RAW WATERIAL ORDER RATE (a)
2000
1750
20. 30. !!lO . 50. 50. 70. Bo. boo ~oo . ~ 10 . ~20 .
TIME (WEEK) SIMULATED TIME
(OMS) ORDER BACKLOG
(WMS) DESIRED BACKLOG
11lMS) RAW MATERIAl. STOCKS
(WMS) DESIRED RAW MATERIAL STOCKS
BASE CASE BEHAVIOUR (b)

Fig.9.3 The DMC model base case.
improvements have been achieved by any of the numerous changes we

might make?
On the face of it, the original problem definition merely called for vari-
ables such as Raw Material Order Rate to be brought under control, but we
can clearly see that there is more to it than that. For instance, Raw Material
Stock matches very badly against Desired Raw Material Stock , although, on
the other hand , Backlog tracks very closely with Desired Backlog . Produc-
tion Rate oscillates quite nicely within the oscillations of New Order Rate,
but it would be desirable to decrease its amplitude a little, as changing
Production Rate is, in the real world, rather an expensive undertaking. If

you have been running the models and studying the output closely you will
have seen that PR oscillates between 83 and 120 WMSIWEEK in the base
case, but the amplitude is from 87 to 117 in Fig. 7.9 and from 93 to 115 if
SWITCH2 is set to 1 (which has the effect of introducing the new policy of
using Raw Material Stock to regulate Production Rate). We thus know that
Production Rate can be reduced in amplitude, and it may be worthwhile to
do so.
This close examination of the output from Fig. 9.3 suggests that there are
four objectives to be achieved:
• to make Raw Material Stock come closer to Desired Raw Material Stock
• to ensure that Backlog does not move too far away from Desired
Backlog
• to avoid unnecessary fluctuations in Production Rate
• to bring Raw Material Order Rate under control.
We have rather arbitrarily put the apparently obvious objective last to
suggest that these four factors are not necessarily of equal importance.
Indeed, common sense suggests that they are manifestly not of equal signifi-
cance to the firm, a fact we shall have to allow for in the final objective
function.
Developing an objective function

To develop an objective function we start by formulating equations to
penalize failure to meet the target factors discussed above .
For the first, we wish to penalize differences between Desired Raw
Material Stock and actual Raw Material Stock , DRMS and RMS, and, by
analogy with the equation formulated in Chapter 8, we must use level
variables to accumulate the discrepancies over the length of the run and we
need to square the differences to avoid positive and negative discrepancies
cancelling each other out. Thus:
L PEN1.K=PEN1.J + DT *(DRMS.J - RMS.J)* *2
N PENl=O
should fit the bill. It should be obvious that it does not matter whether one
writes (DRMS.J - RMS.J) or (RMS.J - DRMS.J); we could equally well
have used EPS2.J. PEN1 simply denotes the first of the four factors which
are to be penalized in the final objective function .
As was mentioned in Chapter 8, the components of an objective function
do not necessarily have to be tested for dimensional validity to the same
extent as those parts of the model which represent the real system. In
this case, PENl has dimensions of [WMS**2*WEEK] , which is meaning-
less, so we shall simply call them Arbitrary Units, ARBUNIT. The same
arbitrary unit notation will be used for the other penalty measures
The DMC problem 255
deduced below, though it will become clear that the arbitrary units for the
last two are [[WMSIWEEK]**2*WEEK] . It does not matter if we lump all
these different units together under the same title of ARBUNIT, as the
components of the objective function are something which we are develop-
ing to help us measure the performance of the model , and they do not
correspond to aspects of the real system.
Similarly,
L PEN2 .K=PEN2.J + DT*(BLOG.J - DBLOGJ)**2
N PEN2=O
should suffice for the variations in backlog.
For these two factors there is a clear target, such as DRMS, against which
the actual, RMS, can be compared (and similarly for the backlog), but there
is no such target for Production Rate. We shall have to invent one, and it is
one of the attractive characteristics of optimization that it frequently re-
quires the modeller to use his imagination and creativity. It is, of course ,
one of the snags of optimization that, if one chooses silly objectives, one will
get silly results.
We cannot use BLEV, the base level for the demand pattern, as a target
for PR, because BLEV is not known within the firm; it is, as it were, known
only to the deity whom these managers have to serve. In any case,
that would gear our optimal policies and structures too heavily to a
stable sinusoidal demand pattern, which would not be very robust.
Instead, we can compare Production Rate to the Average Production
Rate, on the grounds that, if Production Rate does not veer too far from
its own average, relatively fewer production changes will have taken
place and the costs of production changes will have been avoided. This will
give:
L PEN3.K = PEN3 .J + DT *(PRJK - APRJ)**2
N PEN3=O
It would not be reasonable to use Average Order Rate, as we wish to avoid
production variations, not chase order changes. By the same reasoning, we
could penalize differences between Raw Material Order Rate and its own
average, Average Raw Material Order Rate. That is a new variable which
does not exist in the model, so we shall have to add it. The penalty equation
will be:
L PEN4.K = PEN4J + DT *(RMORJK - ARMORJ)**2
N PEN4=O
These four penalties now require to be combined to give an overall objec-
tive function , OF. Since OF will involve the sum of four levels, it has to be
an auxiliary. However, as we have already stated, the four factors to be
penalized are not of equal significance to management, so we will introduce
four weighting factors and get:
A OF.K=(W1*PEN1.K+W2*PEN2.K+W3*PEN3.K+W4*PEN4.K)
As in Chapter 8, it will be neater if OF has some nice round value , such as
100, in the base case, so we introduce a scaling factor to achieve that and
have:
A OF.K=(W1*PEN1.K+ W2*PEN2.K + W3*PEN3.K + W4*PEN4.K)
X/SCALE
where the numerical value of SCALE is found as described in Chapter 8.
When these changes are added to FIG9-3.COS the result is FIG9-6.COS
which should be studied closely, not least to discover the numerical
values for the four weights, WI to W4. Note that all the new variables and
constants have been defined and the new variables which comprise the
objective function have been added to the output commands. Keeping up
with documentation as a model develops is essential good practice. Printing
out the values of new variables is also essential, as much experience shows
that relying on plotted graphs does not give enough information to enable
the modeller to be sure that the revised model is doing the right things for
the right reasons.
Study of the printout for the base case run shows that the penalty compo-
nents of the objective function have significantly different values. In fact,
PEN1 = 16.99 X 106, PEN2 = 50.581 X 104, PEN3 = 7.944 X 103 and PEN4
= 62.111 X 1Q4 at the end of the run , PEN1 actually being about 2000 times
as large as PEN3. Does this matter?
The answer is that it does not . The penalties measure correctly, and in the
correct units, what we have decided to be important in the behaviour of the
system and, if the penalty for raw material stocks has turned out to be very
large indeed, then that is how matters stand. To attempt to scale the penal-
ties so that they were all of the same numerical magnitude might look more
elegant, but would be to second-guess what we have carefully considered to
be the requirements for the system.' In any case, even apparently crude
objective functions tend to produce good results from a model.
Having developed an objective function , the next question is how to use
it. A good way to start is to identify what can be controlled in the system. In
this case, there are two control variables, PR and RMOR, which are, of
course , the domains of the two managers. (There is rarely such a neat
correspondence between control variables and management respon-
sibilities.) For each of the control variables, it is useful to draw up a policy
option diagram, POD. That for PR is shown in Fig. 9.4. This is intended to
show the options which affect PR or which could affect it, which is why it is
1 In the real case on which this textbook example is based a good deal of effort went into
deciding what the system should be achieving, though the insight into what it m ight be capable
of achieving had been generated by the preceding progr amme of experimenting, or playing,
with the model.
The DMC problem 257
called an option diagram. The solid lines show what affects PR in the base
case of the present system. If we ignore BET A for the moment, TABL,
WAO and TAOR are all parameters on the chain which leads to PRo WAO
and TABL are shown at a so-called Policy Level because they regulate the
daily operation of the fixed, currently chosen, control structure shown by
the solid lines.
The firm could, however, adopt a different control strategy of allowing
EPS2 to affect PR , and this is hinted at by the broken lines, to imply
parameters which can only come into play when SWITCH2 is 1. In such a
case, the dotted lines would have to become solid, and EPS2, DRMS and
RMS would move up into the policy level to indicate that TARMS and
1I
Production
Rate
Control PR
Variable
. - - - - - -..........., - - - - - - - - - j - - - - - - - - - r - - ,
1 1 I
_TABL
1 1 I
Backlog 1 I I
Discrepancy
1 I I
Ell 1 1 1
Policy
I I
I I I
Level I I 1
Desired Order 1 I 1
Backlog Backlog
DBLOG BLOG 1 I I
1 1 I
WAO I I I
" " . " • . " " . . . . ..• .. .. ". " " .. ". " t . " " " " . " . " " " t " " "I " "
SWITCH2~ I 1
' - - -TARMS 1 1
R~ 1 I
Mm~1 I I
Stock I 1
Dbcreponcy I
EPS2
I -p----,
1 " I1
---~---I ::
Strategy 1 I I
Level Desired R~ I 1
Row Material 1 I
Mscr~' Stock I 1
---r----
DRMS RMS I 1
1 I I
AveraQ8 J---WAP I
Order 1. Average I ?
Rate --I---pro~:Ion--
· ··· · ···L:. ·······' ····'~~ ···

Exogenous
Level
New
Orde<
Production
Rate
Rate PR
NO
Fig. 9.4 Policy option diagram for production rate .

WAP are now policy parameters for this new strategy. In practice, the effort
of redrawing the diagram would not be worth it. In fact, once one has
understood the concept of a POD and had some practice, it may not be
worth the effort of drawing the diagram at all, except , perhaps, for the
benefit of a client; one simply does it in one's head. It is always a good
idea not to let the conventions of diagrams get in the way of doing the
analysis.
A further variation would be to allow GAMMA to be less than 1, which
opens up the options of to some extent basing DRMS on AOR as well as
the obvious, base case, usage of APR. Similarly, if BETA was allowed to be
less than 1, the dotted line at the right-hand side with 1-BETA pointing to
it would be brought into action.
The? at the extreme right-hand side of Fig. 9.4 suggests that there may be
other structural options not yet conceived of. We shall optimize the model
with the options identified ; the reader should experiment further.
The POD for RMOR is shown in Fig. 9.5, which is to be understood along
the same lines as Fig. 9.4. Note that WAP and TARMS are strategic, or
structural, parameters for PR because they only have an effect if SWITCH2
is 1. In Fig. 9.5, these same parameters are at the Policy Level, because
they apply to the base case. The opposite arguments apply to TABL and
WAO.
The external driving forces are shown as the Exogenous Level, with
TAOR and TAPR as the parameters. For PR , the true exogenous force is
NO . If, however, in Fig. 9.4 SWITCH2 is 1 and GAMMA is larger than 0,
the passive influences from PR to DRMS and hence to EPS2 and PR at the
top of the page would become active and dotted lines would become solid,
as shown in Fig. 9.5. PR has become an exogenous force for PR; in other
words, a new feedback loop has been brought into existence, and this is the
loop shown with emphasized lines in Fig. 9.2. Policy option diagrams are
often quite useful in suggesting feedback loops which one might not other-
wise have thought of.
The practical difference between the Policy and Strategy Levels is that
the policy parameters can usually be changed without too much fuss or
upset in the organization. Any change to a strategy parameter, on the other
hand, usually corresponds to changing someone's way of working, and
that is not always politically desirable in the real world. Fortunately, in
the world of the textbook we do not have to worry about such
considerations.
The first optimization we shall test is the policy level of allowing WAP,
TABL, TAOR, TAPR and TARMS to form the parameter plane. WAO is
excluded for the moment as it represents the basis of the relations between
DMC and its customers and is unlikely to be changed easily. DDEL is
similarly excluded as changing DDEL would involve a complete upheaval
in DMC's dealings with its suppliers; a distinctly non-trivial undertaking.
Each of the selected parameters is allowed to lie within a range from 2 to 10.
The DMC problem 259
Control
Variable
j Raw
Material
Order
Rate
RMOR
r-------- - ---------~-~---~
I Raw
I Material
I Stock
Policy
Level
: I ~T I
I or~ed M~I
I Mat.. ~01 Stock
I Stock RMS
I DRMS
I WAP
I
.. .. . . 1. - .
I SWITC H 1~
I , _TCRMP
I Raw Mateool
I on Order
~ t-c DlscE~ncy a
I I
I - ------,
I Raw ~ateool I
I on Orde< R~~I
Strategy I RMOO ce oroer
Level I DRMOO
I DWRMP---:
I I
I Average I Aver~ ?
1 __ ~:'------r--l- · ~'"'
········G ·· '~ ······ ··· ····· L···· ···

Exogenous \
Level
New
~~:
NO
TAOR Pr"taf"tlon
PR
TAPR
Fig. 9.5 Policy option diagram for raw material order rate .
The precise details of the COSMOS opt imization procedure do not concern
us, though they are fully explained in the COSMOS manual.'
The result of this policy optimization is that, after some 30 iterations, the
values in Table 9.1 are produced.
This rather dramatic fall of more than 90% in the value of OF is mirrored
in the dynamic behaviour with these values, shown in Fig. 9.6A. Note that
the scale has been changed to allow the differences between successive
optimizations to be seen more clearly.
2 COSMOS supports two othe r optimizati on approach es, Base Vect or optimizat ion for iden -
tifying especially sen sitive par am et er s, and Simplification , which can be used to det ermine
which part s of a model ar e really important to its behaviour. They are fully describ ed in the
manu al but are not con sider ed in thi s textb ook as we aim to be as ind ependen t as possibl e of
spe cific softwa re .
Table 9.1 Results of first optimization
Parameter Final value Original value Lower limit Upper limit
WAP 4.00 8.00 2.00 10.00

TABL 10.00 4.00 2.00 10.00
TAOR 8.63 4.00 2.00 10.00
TAPR 10.00 4.00 2.00 10.00
TARMS 10.00 4.00 2.00 10.00
Initial value of objective function 100.00
Final value of objective function 7.78
Figure 9.6A shows that control can be achieved by simple policy changes.
The fact that all the delay parameters, TABL, TAR, TAPR and TARMS,
have been increased in the parameter values shown in Table 9.1 suggests
that the change required is to react more slowly to discrepancies.
The reduction in W AP corresponds to stock targets being set to lower
values.
Four of the parameters have been driven to the rather arbitrary limits
chosen for them, and it may be that using wider limits would produce even
better results. The reader may wish to experiment, but, in practical studies,
one has to give close attention to changes which the people concerned
might be prepared to accept and not lose a good result by pushing them too
far in search of an even better one.
TAOR ends up with the rather strange value of 8.63.There is no particu-
lar significance to that number. How much difference in behaviour would
there be if it was rounded off to 10?
Having achieved a good result at the simple policy level, it is worth
seeking out the effects of policy and structural changes. The results in Table
9.2 are achieved, in which the parameters have been grouped to separate
policy parameters from their structural counterparts.
The objective function is reduced by about 30% and all the policy param-
eters are driven to their extrema, as are some of the structural ones. These
results require some thought to interpret.
That ALPHA and DELTA are driven to 0 implies that the information
structure in the base case is fundamentally misconceived, a point which the
reader should confirm by studying the influence diagram. BETA, however,
is only slightly reduced from the base case value, suggesting that the Pro-
duction Manager should continue to give most weight to Average Order
Rate, but that a little attention to Average Production Rate will also be
helpful.
GAMMA remains at 1.0, implying that it is rather important to keep
stock related to production.
There is, however, a deeper subtlety in these results, which is that the
beneficial effects of the structural changes are dependent on the policy
The DMC problem 261
200
175
150
125
l00+i-+-_u
75
50
25
TIME (MEEK) SIMlA-ATED TIME
(MMS/MEEK) NEll ORDER RATE

(MMS/MEEK) PRODUCTION RATE
(MNS/MEEK) RAM MATERIA~ ORDER RATE
1000
900
BOO
700
600 ++++-_-J
500
100 .
1 L 'BL OG
TIME
(M.S)
(MEEK) SI .....ATED nME
ORDER 8ACK~DG
1,'1 --DBLOG (M.S) DESIRED 8ACK~DG
' ~ ' R MS (MNS) RAM MATERIA~ STOCKS
L-'-DRMS (M.S) DESIRED RAM MATERIA~ STOCKS
A - SIMPLE POLICY OPTIMISATION
Fig. 9.6 Optimal results for the DMC model.

Table 9.2 Results of second optimization
WAP 4.00 8.00 2.00 10.00

TABL 10.00 4.00 2.00 10.00
TAOR 10.00 4.00 2.00 10.00
TAPR 10.00 4.00 2.00 10.00
TARMS 10.00 4.00 2.00 10.00
BETA 0.82 1.00 0.00 1.00
GAMMA 1.00 1.00 0.00 1.00
ALPHA 0.00 1.00 0.00 1.00
DELTA 0.00 1.00 0.00 1.00
changes, and vice versa. Put another way, it will be necessary to change
policies to get the best results from the reorganization of structure, and it
will be necessary to reorganize to get the best results from new policies.
This is a most fundamental insight into management theory, which helps to
explain why corporate reorganizations often fail and why policy changes
sometimes fail to have noticeable effects.
The dynamics of the second optimization are shown in Fig. 9.6B. Com-
pared with Fig. 9.6A, the control of Raw Material Stock is rather better and
that of Backlog is a little worse. Raw Material Order Rate is shifted in time
so that it is closer to the cycles in Production Rate. Close study of the
printed output shows that Production Rate oscillated between 84 and 119 in
Fig. 9.6A and between 86 and 118 in the second case. That difference is hard
to detect from the plotted graphs, so close study of printed output is always
useful in detecting these second-order effects.
It sometimes happens in real analysis that powerful interests or personal-
ities are simply unwilling to countenance changes, no matter how beneficial
those changes might be to the wider organization. To illustrate that concept,
the next optimization does not use TABL and DELTA and produces the
results shown in Table 9.3.
Surprisingly, this is actually slightly better than the previous case. Again,
the fundamental insight is that the policies and structures required are quite
different when some options are ruled out. The parameter plane and the
response surface have changed considerably, but optimization has found
the best that can be done in these different circumstances.
The graph in Fig. 9.6C is also significantly changed from Fig. 9.6B. The
control of Raw Material Stock is better, as is that of Backlog. The oscil-
lations in the two backlog variables are much steeper and match more
closely, and the amplitude of Production Rate is now from 81 to 121.
Studying the dynamics is an essential aspect of interpreting optimization
200
175
150
125
100+1-+---'-"
75
50
25
TI ME (WEEK) SIMlA-ATED TIME

IWWS/WEEK) NEW ORDER RATE
l:=~~:~~l :~Ol~mg~A~A ~~ER RATE
1000
900
BOO
700
»:
I
600 +ti+-_.F
500
400
300
200
100
IIIL' TI ME (WEEK) SIMlA.ATEO TIME
I'I:--DBLOG
BLOG (WWS)
(WWS)
ORDER BACKLOG
DESIRED BACI<LOG
' ~ ' A MS (WMS) RAW MATERIAL STOCKS
L.'-OAMS (WWS) DESIRED RAW MATERIAL STOCKS
B - POLICY ANO INFORMATION STRUCTURE OPTIMISED
Fig. 9.6 Continu ed

Table 9.3 Optimization without TABL and DELTA
Parameter Final value Original value L ower limit Upper limit
WAP 4.00 8.00 2.00 10.00

TAOR 2.87 4.00 2.00 10.00
TAPR 10.00 4.00 2.00 10.00
TARMS 10.00 4.00 2.00 10.00
BETA 0.77 1.00 0.00 1.00
GAMMA 1.00 1.00 0.00 1.00
ALPHA 0.00 1.00 0.00 1.00
In itial value of objective function 100.00
Table 9.4 Optimizing structures in isolation
Param eter Final value Original value L ower limit Upper limit
BETA 0.66 1.00 0.00 1.00

GAMMA 1.00 1.00 0.00 1.00
ALPHA 0.00 1.00 0.00 1.00
DELTA 0.00 1.00 0.00 1.00
SWITCH1 1.00 0.00 0.00 1.00
SWITCH2 1.00 0.00 0.00 1.00
output. The objective function value for this third case is better than it was
for the second, but whether the behaviour is really better might be a rather
moot point. The objective function should be understood for what it is: a
mathematical artefact to guide the optimization search, not an arbiter of
what is good and what is not. The ultimate choice of whether or not to
accept and implement results derived from optimization should be made by
reviewing the output carefully. In practical cases, reviewing the output
might make one decide that the components of the objective function and!
or their weighting factors did not adequately reflect the objectives
of the problem proprietors. For that to happen is usually good, as the
optimization has helped to clarify understanding, as was suggested in
Fig. 1.4.
Having looked at policies on their own, it might be illuminating to exam-
ine the optimization of structures in isolation, the results being shown in
Table 9.4.
The objective function is improved still further, and the pattern of earlier
results tends to be confirmed in that ALPHA and DELTA are forced to 0,
200
175
150
125
100 -1-:-+-._.1.:.
75
50
25
TIME lWEEKI SIMULATED TIME

IWMS/WEEKI NEW ORDER RATE
IWMS/WEEK) PAOllUCTION RATE
IWMS/WEEK) RAW MATERIAL ORDER RATE
1000
BOO
500
L'
TI ME (WEEK) SIMULATED TIME
'II BLOG 1_' ORDER 8ACKL06
. _..:::: ~a§OG IWNS) DESIRED BACKLOll
1-) RAw MATERIAL STOCKS
l -'-ORMS (_I DESIRED RAW MATERIAL ST0Cl<8
C - REFUSAL TO CHANGE TABL AND DELTA

Fig. 9.6 Continued
Table 9.S Optimization using all policy and structural parameters
WAP 4.00 8.00 2.00 10.00

TABL 10.00 4.00 2.00 10.00
TAOR 8.68 4.00 2.00 10.00
TAPR 10.00 4.00 2.00 10.00
TARMS 10.00 4.00 2.00 10.00
BETA 0.76 1.00 0.00 1.00
GAMMA 1.00 1.00 0.00 1.00
ALPHA 0.00 1.00 0.00 1.00
DELTA 0.01 1.00 0.00 1.00
SWITCHI 1.00 0.00 0.00 1.00
SWITCH2 1.00 0.00 0.00 1.00
GAMMA remains at 1 and BETA is still weighted towards the use of AOR
in the Production Rate decision. Interestingly, both SWITCH1 and
SWITCH2 reach their limits of 1, suggesting that the MIS should be bought
and that the raw material stock discrepancy should be used in the Produc-
tion Rate decision. As the MIS is expensive, it would be interesting to run
the model with the above parameters, apart from keeping SWITCH1 at 0,
to see how much difference there is. The dynamics of this solution are to be
seen in Fig. 9.6D, and it is now time for the reader to start assessing the
quality of that performance against earlier solutions.
The next set of results (Table 9.5) are obtained when all the policy and
structural parameters are used in the parameter plane.
There is a further improvement in the objective function and the pattern
of results is maintained. The dynamics appear in Fig. 9.6E.
Finally, we optimize the model with the addition of TCRMP to the
parameter plane. This is to drive home the idea that , if the MIS is im-
plemented when SWITCH 1 is 1 - a structural change - it is essential to
examine the policy parameter(s) which might enable the best to be
achieved with the changed structure. This gives results, shown in Table 9.6,
which are consistent with the patterns achieved earlier, in the sense that
most of the parameters are forced to their upper extrema. The exception is
TCRMP, which is forced to its lower limit. In essence this means that the
company should take a generally relaxed attitude to discrepancies, but,
having decided to buy the MIS, the attitude should be to use it aggressively
to manage the amounts in the pipeline. However , to enable those policies to
have their full effect, changes to information structures will be needed. At
the risk of repetition, this idea that policy and structure need to be harmo-
niously tuned is quite fundamental although , without the power of the
200
175
150
125
100 +r-+----"-
75
50
25
TIME (OEEK) SIMIA.ATED TIME
10MS/OEEK) NEO ORDER RATE

'OMS/OEEK) PRODUCTION RATE
(OMS/OEEK) RAO MATERI4L ORDER RATE
1000
900
BOO
100 .
o . 40. 50 . 50 . 70 . eo. 90 . ~oo . ~ 10. ~20 .
L'
',"I BLOG
--OBLOG
TI ME
(OM9)
10MS)
(OEEK) SIMlLATED TIME
ORDER BACKLOG
DESIRED BACKLOG
' ~ ' RMS 10MS) RU MATERI4L STOCKS
' - '-DRMS (OM9) DESIRED RAO MATERI4L STOCKS
o - STRUCTURAL CHANGES
Fig. 9.6 Continued
200
175
150
125
100~-l-_J.:
75
50
25
TI ME (WEEK) SIMlA.ATED TIME

(WMS/WEEK) NEW ORDER RATE
(_/WEEK) PROllUCTIOM RATE
(_/WEEK) RAW MATERIAL ORDER RATE
1000
900
BOO
500
400
100 .
I11!L'~LOG TI ME (WEEK) SIIU.ATED TIME

(WMS) ClADER IIIoCKL06
.I ~~~B!;OG (WMSl
(WMS)
DESIRED BACKLOG
RAW MATERIAl. STOCKS
L-'-ORMS (WMS) DESIRED RAW MATERIAL STOCKS
E - STRATEGY AND POLICY
Fig. 9.6 Continued

The DMC problem 269
Table 9.6 Optimization using TCRMP
Param eter Final value Original value L ower limit Upper limit
WAP 4.00 8.00 2.00 10.00

TABL 9.99 4.00 2.00 10.00
TAOR 5.90 4.00 2.00 10.00
TAPR 10.00 4.00 2.00 10.00
TCRMP 2.00 4.00 2.00 10.00
TARMS 10.00 4.00 2.00 10.00
BETA 0.77 1.00 0.00 1.00
GAMMA 1.00 1.00 0.00 1.00
ALPHA 0.00 1.00 0.00 1.00
DELTA 0.00 1.00 0.00 1.00
SWITCHl 1.00 0.00 0.00 1.00
SWITCH2 1.00 0.00 0.00 1.00
optimizer to search the options, it would run the risk of remaining little
more than a pious intention.
The dynamics of this solution are shown in Fig. 9.6F. This is by far the
best value so far found for the objective function , but you should form your
own assessment of the dynamics shown in Fig. 9.6F relativ e to some of the
previous cases. In fact, we have reached a stage at which you should pause
and review all the opt imizations so far done for DMC.
Summary of optimization of DMC

So far, we have discussed in some detail the approach to optimization,
illustrating the method with the DMC problem. Before proceeding to assess
the value of optimization, we must review the process.
At first glance , the idea of optimizing something is so attractive that a
naive analyst might jump to the conclusion that all one has to do is switch
on the software and all will be revealed. However, we have tried to show
that optimization is not quite so simple as that, and we have worked
through several steps.
• After studying the base case performance of the model we decided on the
factors to be aimed at; four in all. We were led to these four factors,
rather than the obvious lone factor of achieving control of raw material
ordering, by the experiments we had performed in Chapter 7 in which we
explored the range of behaviour trade-offs which this model possesses . In
a practical study, this would have involved much discussion with manage-
ment. Interactive software and notebook computers are a great help in
those consultations. Deciding what the system is supposed to achieve is,
200
175
150
125
100+1--!-_.Jv
75
50
25
TIME (MEEK) SIMlUTEO nNE

(MMa/MEEKI NEM ORDER RATE
(MMa/MEEKI PROoocnON RATE
(MMa/MEEKI RAM MA TERIA~ ORDER RATE
1000
900
800
700
6004-++1---/
500
400
300
200
100
TIME (MEEKI SIMULATEO nME
I_I ORDER 8ACK~OG

(_I DESIRfIl 8ACK~06
(MMa) RAM MA TERI~ STOCKS
(MMaI DESIRfIl RAM MA TERI~ STOCKS
F - ADDITION OF TCRMP TO RUN E
Fig. 9.6 Continued

however, the crucial first step . We shall have much more to say about that
at the end of the chapter, the point at which we shall explain why simply
minimizing cost or maximizing profit are not usually suitable as
objectives.
• Having decided on the objectives, the next stage was to formulate a
measure of how well they are achieved: the objective function.
• Armed with the objective function, we next considered how to use it and
suggested the policy option diagram as a pictorial support to thinking
about the design options available . Again , we were led to those options
by the previous work of experimenting, or playing, with the model. It
should be clear by now that sensible optimization cannot take place with-
out prior experiments.
• Even with the objective function and the options the optimization pro-
ceeded in several stages; this is simulation by repeated optimization at
work! At each stage , the value of the objective function improved, but
the dynamic behaviour did not necessarily improve to the same extent
when it was studied by eye. Just as the eye enabled us to formulate the
objective function, the eye should evaluate its results. In practical studies,
visually evaluating optimization results often leads to an improved
objective function. This is no great burden, as optimization is so
computationally efficient in COSMOS that the extra cost is trivial. The
real point is, however, that the objective function is not the arbiter
of performance; that role belongs to human judgement after the
objective function has fulfilled its role of controlling the hill-climbing
computations.
This summary of the process does, however , beg the question of whether
optimization is really of value. Not to put too fine a point on it, is opti-
mization through several stages worth the effort and does optimization
produce better results than the experimentation which must necessarily
precede it? It is dangerous to generalize from a small sample , but we
shall discuss the question, using DMC as an example, in the next
section.
ASSESSING THE VALUE OF OPTIMIZAnON
There are three bases on which we ought to weigh the value that opti-
mization adds to system dynamics. The first is whether the last optimization,
Fig. 9.6F,3 is genuinely better than the first (Fig. 9.6A). In other words, has
the study of policy, strategy and structural options taken us much further
than simply allowing the software to search the obvious parameters? The
second is whether the best optimization result is better than the 'best'
JThroughout this section , when we refer to Fig. such-and -such we mean the graph of the
dynamic behaviour and the optimizati on results which accompan y it.
experiment. The third is whether the simplest optimization is better than

the 'best' experiment.
Notice that we have referred to the 'best' experiment in quotation marks.
As we said at the beginning of this chapter, the snag with the experimental
approach is that one never knows whether just one more experiment would
have turned up something even better, and one can become rather old, and
practical studies can turn out to be very expensive, if one continues to try
just one more time. Optimization offers the prospect of much more power -
ful searches - the optimizations we have done in this chapter take literally
a few seconds on a 486 DX 2 PC running at 33 Megahertz - but how good
are the results?
To answer the first question of assessing the value gained by repeated
optimization, we need to contrast Fig. 9.6F's results with those of Fig. 9.6A,
which gave a final objective function value of 7.78. On the face of it,
Fig. 9.6F is more than 7 times better than Fig. 9.6A because its final
objective function value of 1.07 is that many times as good as the 7.78 which
was achieved in Fig. 9.6A. When the graphs are compared visually, how-
ever, it is very hard to say that 9.6F is many times better than 9.6A. It is,
however, a very definite improvement, so the successive optimizations have
served their purpose in guiding us to a better solution; we must simply be
careful not to overstate matters and say that one solution is x times as good
as another simply on the basis of the ratio of the two objective function
values.
Why, however, did we bother about all these optimization stages? Why
not simply allow the parameter plane to consist of all the options identified
in the policy option diagram and let the software pick out the best solution?
There are two answers.
The first answer is that the simplest change to a system is always the
easiest to get accepted. People resist upheaval, so disturbance ought to be
minimized, and a minimal alteration to established habits is much more
likely to be accepted by the proprietors of a problem. This is true in fee-
earning studies for organizational clients and it is also true for academic
research aimed at policy recommendations. If, in this instance , the differ-
ence between Fig. 9.6A and Fig. 9.6F had been trivial, then the simple
changes to base case policies would have been all that was necessary. There
might, of course , be a mid-way point at which the improvement is good
enough to be acceptable and the changes involved are small enough to be
tolerable.
The second answer is that optimization is not a substitute for thought,
and there is usually much to be learned about the system from simulation
by repeated optimization just as there is from successive simulation
experiments, as we saw in Chapter 7. In this instance , Fig. 9.6F includes the
result SWITCHl=l; in other words, a recommendation to purchase the
Raw Materials Manager's MIS. The experiments in Chapter 7 showed cases
in which the MIS added little or nothing to performance, so the reader

should re-optimize Fig. 9.6F without allowing SWITCHI to lie in the pa-
rameter plane. The final value of the objective function mayor may not be
worse and the result ing parameter values mayor may not be the same as
Fig.9.6F.
To see whether or not the final optimization is better than the 'best'
experiment, the second question, consider Fig. 9.7 (FIG9-7.COS on the
disk). That has been plotted to compare Fig. 7.9 with Fig. 9.6F, and the
variables shown are the Production Rate and the two error terms, EPSI
and EPS2. Interestingly, the oscillations in PR are slightly smaller in
Fig. 7.9, but the behaviour of EPSI and EPS2 is significantly improved .
It is very likely that someone who has a deep intuitive understanding of
system behaviour, or who spends enough time experimenting, will come to
results which are as good as optimization can achieve, but normally experi-
mentation teaches enough about the properties of the system for optimiza-
tion to be let loose in the profitable directions so as to find still better
results. Even if the results are no better than the experiments, which is
unlikely, one's confidence that they are the best that can be achieved is
greatly increased.
Our third question was whether simple optimization, confined to the
obvious parameters of the base case, is better that the 'best' experiment. To
see that, consider Fig. 9.8 (FIG9-8.COS on the disk), which compares Fig.
7.9 with Fig. 9.6A. For all practical proposes there is no difference whatever
between the two sets of graphs.
For this model, we can conclude that:
• A programme of successive optimizations does payoff and produces

better understanding of the options available than simply choosing a set
of parameters and letting the software loose on them.
• Optimization improves on, and adds confidence to, experimentation
when the results of the experiments are used to provide a basis of insight
for optimization .
• Optimization does not help very much when it is performed without
thought, using obvious candidates for the parameter plane.
This all seems very much like common sense, which is comforting. It is
typical of what has been found in numerous uses of optimization but,
just as hill-climbing optimization is based on a heuristic rule of thumb ,
different patterns of result are sometimes found . A truly talented, or very
lucky, experimenter may avoid getting stuck on a subordinate peak of
the response surface, which can sometimes happen with optimization.
Needless to say, a simple-minded objective function is likely to produce
simple-minded conclusions, a point we shall review at the end of this
chapter.
250
225 I
11
200
175
150
125
100-H __
::L~
25 I
0: . ~~cr ~o. 50. so. 70. Bo . 90 . 1100 . I,10 . ~20.

:L TIME (MEEK) SIMULATED TIME
i-- PR
PR
' BEST ' EXPERIMENTAL RESULT FROM FIG 7 .9
OPTIMAL RESULT FROM FIG 9 .8F
(a) PR (WMS /WEEK) PRODUCTION RATE

100
75
50
25
o
- 25
- 50
-7 5
L. 50. so. 70. Bo . 90 .

-100
1tl~30 . ~o . 1100 . 1'10 . ~20 .
I TIME ("lEEK) SIMLLATED TIME
!-- EPS1
EPS 1
' BEST ' EXPERIMENTAL REst.LT FROM FIG 7 ,9
OPTIMAL RESUL.T FROM FIG 9 .6F
(b) EPS1 (WMS) BACKLOG ERROR

100
75
50
25
-25
-50
- 75
:L. ----;1"'0'-.-!<""o"-. -k Bo. eo. go .

-100
~o . 50. 70. 1100 . 110. ~20 .
TIME (MEEK) SINULATED TINE
: EPS2 ' BEST' EXPERIMENTAL RESULT FROM FIG 7 . 9
(c) -- EPS2 OPTINAL RESULT FROM F I G 9 .8F
Fig. 9.7 Experimentation versus optimization.

250
225
200
175
150
125
1oo-i-+_ _~
75
50
25
. .-J.rn-:- 50. 60. 70. 130 . 90. 1100 . 110. ~ 20.

o -'-+.r---I,-,.r--lom----o,,.,--
TIME (lI£EK) SIIU.ATED TIME
' BEST' EKPEAIMENTAL AESlA.T FROM FI6 7 .6
SIMPLE OPTIMAl.. RESLl.T FROM FIG SI .e"
(WMS /WEEK) PRODUCTION RATE (a)
75
50
25
-25
-50
-75
-100
. - -1m ~ -70-:- !m:-lro. 50. 60. 70. 130 . 90. 1100. 1'10 . ~20.
I L TIME (IIIEEK) SIMULATED TIME
J EPS1 ' BEST ' EXPERIMENTAL RESULT FRON FIS 1 . g

- -EPS1 SIMPLE OPTIMAl. RESu, T FRON FI6 9 .e'
EPS1 (WMS) BACKLOG ERROR (b)

100
75
11
50
25
o
-25
-50
-75
-100
1100. 110. ~20.
TIME (MEEK) SIIU.ATED TIME
' BEST ' EXPERIMENTAL FESlI..T FROM FI8 7 ,g
SIMPLE OPTIMAL RESU,..T FROM FIG g.eA
EPS2 (WMS) RAW MATERIAL ERROR (c)
Fig. 9.8 Simplistic optimization.

CONSTRAINED OPTIMIZAnON
So far, we have examined an optimization problem in which we were free to

choose whatever parameters we wished for the parameter plane and the
parameters could be changed without incurring any expense. We now need
to consider the case of constrained optimization, in which money (or some
other resource, such as people) has to be spent to make changes to the
system. To do so, we shall use the simple combat model as an illustration
(this is an exceedingly simple model and the reader should not think it is
typical of system dynamics models developed for real defence studies!).
System dynamics has been used very successfully in defence problems, and
the methods for constrained optimization were first developed in that prob-
lem domain, though they have been equally successful in other problem
areas .
The distinction between free and constrained optimization is, however,
not absolute. Most constrained problems contain at least some free
parameters, and the combat model is no exception. In particular, the
parameter Blue Commander's Acceptable Force Ratio can be changed
freely if he can be persuaded to change his mind about his intentions
for fighting the battle. In a business strategy problem, the constraint is
often on the amount of money which the firm can invest, but there are
usually free parameters such as the ratio of long-term to short-term debt
which management judges to be prudent. As we shall see below, the free
and constrained parameters interact in ways which produce quite profound
insights.
There is another distinction between the two forms of optimization,
which is the shape of their respective response surfaces. In unconstrained
optimization, the response surface is best envisaged as a mountain range , as
we sought to indicate in Fig. 8.3. In constrained optimization, however, it is
evident that, given enough resources, performance can be increased practi-
cally without limit. The snag is that there are limits, and a cost barrier will
be encountered, which is why the optimization is constrained.
This is suggested in Fig. 9.9, which is a perspective sketch of an object
viewed from behind and looking slightly to the right. The two parameters
are in the horizontal plane; the performance index is vertical. From the
origin, at which no money has been spent, the dotted lines indicate a slope
of performance leading off into the three-dimensional distance, though it is
not a smooth slope and contains many wrinkles and troughs. Instead of
climbing a mountain, the blind man now feels his way up the most promis-
ing ridge. He has to pay to advance and, if he makes enough steps , he
eventually encounters a cost barrier when his money runs out.
If, of course , increasing Parameter 1 paid off better than Parameter 2,
then larger increases in the first would have to be paid for by smaller
increases in the second, and the cost barrier should be imagined as
being capable of rotating to favour the most beneficial parameter while
Optimization of the combat model 277
Cost Barrier on the Response Slope
~
~ k~~
\ §§ I
\ §~ I
\ -&! ;/
\ 0
0.. 8 I
'~
{;;~
Q>"
I I
\
'J;;
lJ
g0 I
\ Q) If!; I
, g ~§ I
\ §
, 0 I
I
\ 't: /
, Q) I
,0-
c;
I
I I
i"'CS) \ 1/
%'
"6 \ I
I
0- II
t Origin of Response Slope
~hen there has been no
Investment
Fig. 9.9 Response slope for constrained optimization.
preventing total investment from exceeding the available funds, or what-

ever resource is being invested.
Having discussed these ideas, we shall now apply them to the simple
combat model.
OPTIMIZATION OF THE COMBAT MODEL
For convenience of reference, the influence diagram for the combat model
and its base case dynamics are shown in Figs. 9.10 and 9.11. Figure 9.10 also
emphasizes the 'pressure points'; parameters which could be changed to
alter the system performance, though usually only by spending money or
investing in some way. Obviously, all the parameters in the DMC problem
could be called pressure points, and one of them, SWITCH1, very definitely
involves spending money, but it is useful to reserve the term 'pressure point'
to emphasize in one 's mind that the problem is one of constrained
optimization.
As in the DMC case, the first step is to decide what the system, in this case
the Blue government, is trying to achieve. The base case dynamics show a
disastrous Blue defeat, with Blue's forces completely annihilated by
TIME=33. A simple objective would be for Blue to survive as long as
possible, but that implies an acceptance of eventual defeat. Further thought
leads to the notion that Blue 's forces should survive as much as possible,
which is tantamount to moving the dynamics of Blue Forces in Contact
- -- --
+
, -,
Forc e Raho .....
\ - - (Red / Blue)
,,
\ -- • t- \
\
\ ,,"
\ " ," I\ Il.dllo,.
\ '" @ _-- 01_
\ +/ / \
". ,,-,Man
\, .. I I
\ ~
rcece ac ne
I Triggat'
"" I
I
Tot al Red ..l
/ - Flnr.g Rate - ....
I \ I I ' \
I " " \ I / \
I \
\ I I \
I
I \ ~~ +
I \ Blue 1,05$
\ Rate
-c
1-
\
\
I \
/ \
/ \
/ .... --il- - ,
,,
I /
,,
I
I
'\
I I \
,
+ ,
I Blue Peserv es - Blue Reserv e 81ue in
B1ueRese rve- - . . Combat
Ramoining ...- - - Com m itm ent Rat e B • A mvoi Rote ~
\ ®+
\ • +", t I
.... / \
\ ~ ,,'" / I \
' I _ / "
\
\
\
, +
.... lo tal Blue
'-r-c:~ , ,, Fwing Rot e
,, ,r l
, ,
t-;
'''I/io/ F_,*
'-_
Foe.
~
,
,-
~
,t
,
,,
70101
- - - FoIc. - ~
Fig.9.10 Blue's pressure points in the simple combat model.

60 '11
I
50.
1
I
I
1
I
40 .'
I
1 _------------------ -
30 . : """"",,""---
1 "
20. 1 /
1 /
I ---', ///
10 . 1 ......... _ ........
I
1
1
o..Z-I-+r---l.-;..,..-J....-~k-l".....--I,,,,....___l,,,...:>__l,...___l,,.,.,_-l;,,.,--_t,.,.,.__6,__{,
111031 . O.
L~~~HE
TIME (HOUR) SIMIA.AlED TIME

(RED) REO FORCES IN CONTACT WITH BLUE (a)
2'1 60 '
1 .7~
TI
50·4---,,
1 .~ ,,
,,
1.2J 40. ,,
,,
1 ,
1'1 30.
1\---------------
.7~ 1\,
20 . * -\. \
.~ I \,
I '\
.2~
10 .
\,, I
J ,
0 '~-n-r-4rrr-~-l.:Jrr-'ll:n::---lor""'!=---\"'~=--b,.---l=--{'
O.
I 1110\' . TIME o.
!
(HOUR) 51WULAlEO TIME
BLRES (BLUE) BLUE RESERVES REMAINING

1 -- ARES (REO) REO RESERVES REMAINING
----FRTRIG (1) SNITCH TO TRIGGER CO*ITIENT OF BLUE RESERVES
BASE CASE (b)
Fig. 9.11 Base case for the combat model.
With Red, BLUE,4 as far along the time axis as possible and as far up its
own axis as possible; maximizing the area under the curve, in fact. That
suggests a Blue Performance Index, BPI, as shown below. It is obvious that
BPI is something to be maximized, compared with INDEX in the DMC
case, which was something to be minimized. COSMOS can minimize or
maximize at the user's choice.
4 In this section it will be useful to remember the distinction that 'Blue' means one of the sides
in the conflict, while BLUE is the name of a variable in the model. Review FIG6-10 .COS on
the disk.
note
note sector for objective function
note ===============
note
I bpLk=bpLj+dt*blue.j**2
n bpi=O
In this case, there is no actual need to square BLUE.J to prevent it going
negative; the reason for doing so is partly to magnify the effect of
larger Blue forces surviving, and partly to get into the habit of using
squared values to avoid positive and negative effects cancelling each other.
Acquiring the habits of good practice is an important part of successful
modelling .
Alternatively, one might wish to base BPI simply on the numbers of Blue
who survive at the end of the conflict, disregarding the dynamics of BLUE
during the run . That would require:
a bpi.k = blue.k
In this case, there is no need for BPI to be a level, because we do not wish
to accumulate a memory of performance during the run, only to store the
value at the end of the iteration. The reader should experiment with this
alternative approach.
BPI is only part of the story, because it is obvious that, if Blue spends
enough money on increasing the size of its army, it is bound to reach a stage
where it will win. Unfortunately, Blue does not have endless money , and
the trick of constrained optimization is to subtract some very large number
from the performance index whenever the optimization search tries to use
a collection of parameters the cost of which would exceed the given budget.
In this case, Blue has a budget of $24000 (mythical money!), so we subtract
a huge penalty from BPI whenever that limit is exceeded:
a of.k = bpi.k/scale -1 Oe06*(max(O,cost - budget))
c budget=24000
c scale=20.5846e06
BPI is scaled to 100 for the base case.
Subtracting large penalties in this way is likely to make OF become large
and negative in some iterations, which is tantamount to the blind man
falling off a cliff. The software recovers from that misfortune, rather as
cartoon characters can reverse in mid-air. It can be quite entertaining to
watch the iterations!
How, though, are we to calculate COST?
The key idea is to understand that the software will search among the
chosen set of parameters. Thus, we might allow the initial value of Blue in
Contact With Red, IBLUE, to be in the parameter plane, in which case the
hill-climbing will vary that parameter as it seeks to maximize BPI. If
IBLUE is increased, money has to be spent. If IBLUE is decreased, money
is made available to spend on other things. We need therefore to freeze the
base case value of IBLUE, calling it, say, BIBLUE, so that the financial
consequences of any changes to BLUE can be recorded. This is shown in
italic type in the following equations:
note
note blue forces in combat
note
I blue.k=blue.j+dt*(brar.jk-blr.jk)
n blue=iblue
c iblue=10000
c biblue=10000
This allows us to calculate one component of the cost effects as:
n cost=(iblue- biblue)*costreg
c costreg=1.2
A new value for IBLUE is chosen at the start of each iteration, so the cost
will be fixed throughout that iteration and not vary during it. For that
reason, we can use a computed constant to calculate COST from the values
of other constants.
COSTREG [$/BMEN] is the cost of adding a soldier to Blue's standing
army or the saving if one is eliminated. In practical cases, great care has to
be devoted to ensuring that these costs are correctly calculated. We shall
assume, for simplicity, that the marginal cost and the marginal saving are
equal. Changing the equation to represent a case in which they are not
equal is an exercise in the problems in Appendix B.
FIG9-12.COS on the disk is set up to allow us to optimize with five of the
costly parameters, which requires a cost equation:
n cost=(iblue- biblue)*costreg + (iblres- biblres)*costres
x +(bbmdel-bmdel)*delcst+(bmcap-bbmcap)*capcst
x +(brmcap-rmcap)*rmccst
c costreg=1.2
c costres-su.S'
c delcst=5000
c capcst=6
c rmccst=9
Finally, we must choose some optimization ranges for the chosen pa-
rameters, as shown below.
5 For simplicity we shall assume that the reserve troops are less expensive than those manning
the frontier because it is easier to supply them with food when they are nearer their home
base. We shall, however, suppose that they are as effective as the frontier force. If one wished
to widen the scope of the model, one could assume, instead, that reserves are both less
expensive and less effective than regulars . The reader may care to tackle that as a modelling
project.
note optimization ranges are :

note
note iblue 5000-15000
note iblres 5000-30000
note bmcap 4000-6000
note bmdel 3-6
note rmcap 1500-2500
These ranges for five selected parameters allow IBLUE and IBLRES to
vary independently; that is, they do not consider the pressure point of Total
Force, which was shown in Fig. 9.10. In short, there may be more than one
constraint in a problem; in this case it might be total money and total men.
Amending the model to allow for two constraints is an exercise in Appendix
B.
It should go without saying that you should, if possible, do some
optimizations of your own, using different parameters and inventing your
own cost values.
The necessary definitions are shown below.
d bbmcap=(blue/hour) base value for blue movement capacity
d bbmdel=(hour) base value for blue movement delay
d biblres=(blue) base value for initial blue reserves
d biblue=(blue) base value for initial blue forces in contact with red
d bpi=(arbunit) blue performance index to measure survival of blue's
x forces
d brmcap=(red/hour) base value for red movement capacity
d budget=($) defence spending budget
d capcst=($/(blue/hour» unit cost of increasing blue transport capacity
d cost=($) total cost of set of improvements being considered
d costreg=($/blue) cost of additional regular soldier for blue
d costres=($/blue) cost of additional reserve soldier for blue
d delcst=($/hour) unit cost of reducing blue movement delay
d iblres= (blue) initial blue reserves
d of=(arbunit) blue's objective function to be maximized
d rmccst=($/(red/hour» unit cost of reducing red's movement capacity
d scale=(l) scaling factor to make objective function=100 for base
x case
There are two options for optimizing the model: using only the five costly
pressure points and using the five points plus the free parameter of the Blue
Commander's Acceptable Force Ratio: the degree to which he is prepared
to be outnumbered before committing reserves. It is, of course , a necessary
check to add the computed constant, COST, to the output and check that
the limit has not been exceeded. With the costly parameters, COST ends up
at $23954; in the other case it is $23999.
Table 9.7 Table of final parameter values for the combat model
Parameter Base case Costly parameters Costly and free parameters
BMDEL 6.00 5.06 5.60

BMCAP 4000 5644 5524
RMCAP 2500 1978 2065
IBLUE 10000 7260 10800
IBLRES 20000 30000 30000
BCAFR 2.00 2.00 1.14
These two cases produce rather different results for the two parameter
sets, as shown in Table 9.7 and in Figs. 9.12A and 9.12B. In both cases, there
is a resounding Blue victory, Red ending up by marching obligingly on to
the battlefield to be destroyed by Blue's overwhelming firepower. In Fig.
9.l2A, BLUE's final strength is 30980 and in Fig. 9.12B it is 35940. The
commander's change of mind saves 5000 casualties!
The interesting insight is, however, in Table 9.7, showing final parameter
values. In both cases, IBLRES is driven to its maximum value, because
reserve troops are cheaper than the regulars who form the initial Blue Force
in Contact with Red. However, the second optimization encourages the
Blue commander to commit his reserves at an earlier stage , in which case it
also finds that it will pay him to have more regular troops available at the
beginning, even though he has less money to spend on reducing his move-
ment delay and increasing his movement capacity. In the first case, in which
he is unwilling to change his intentions, it pays him to reduce delay and
increase transport capacity so as to be more easily able to make use of his
increased number of reserves. At the risk of repetition, this is a very simple
model and the results are neither realistic nor typical of those produced by
more serious models.
It is important to try to explain optimization results by such verbal
reasoning, rather than simply reading off tables of numbers from the
optimizer printout.
Apart from explaining the results in verbal terms, the table repays a little
more scrutiny. The key point is that the results for the costly parameters are
different in the two cases of fixing and changing the free parameter. In policy
terms, this is a most fundamental insight, regardless of whether the problem
domain is military, corporate, or any other, strategy. The reason for the
difference is that the change ofpolicy is only fully effective ifthe distribution
of costly assets is changed. Correspondingly, redistributing assets only be-
comes effective ifpolicies are changed. Put another way, being flexible about
policies makes it possible to use assets to the greatest effect, while acquiring
the right assets makes it possible to adopt policies which will have the
greatest effect. This is, of course , remarkably obvious after one has thought
6°·~i
50 . 1
I
I
40 .
30.
20 .
10.
ao . as. 50. 5s . bo o
L
*1~ ·~1k-.-r-,.,.--l,;"""--"""""~'O.-~5.-,J31r ~ .
!--REO
BLUE (BLUE) BLUE FORCES IN CONTACT NITH RED
(RED) RED FORCES IN CONTACT NITH BLUE
60 .
---,,
,,
\
\
\
\
\
\
\
30 ·-i+---'--~7--------------
\
\
\ ,,
\
\
\ ,
\
\
\
\
\ ,,
\
TI ME (HOUR) SI ...... ATED TIME

111 SNITCH TO TRIGGER COMMITMENT OF BLUE RESERVES
A - CONS TRA INED OPTIMISA TION
Fig. 9.12 Opt imizing the combat model.

60 .
50.
40 .
30 .
20 .
10 .-...r--,--
TIME (HOUR) SII4IA.ATEO TIME
(BLUE) BLUE FORCES IN CONTACT NITH RED

(REO) REO FORCES IN CONTACT NITH BLUE
60 .
- - -\
\
\
\
\
\
\
\
\
\
\
30'-i+---r--~_--------------
\
\
\
\
\
\
\
\
\
\
\
\
\
\
\
\
\
\
TIME (HOUR) SI ......ATEO TIME
(BLUE) BLUE RESERYES REMAINING

(REO) REO RESERYES REMAINING
(1) SNITCIl TO TRIGGER COMMITMENT OF BLUE RESERYES
B - CONSTRAINED AND FREE OPTIMISATION

Fig. 9.12 Continued
of it. The role of optimization is to be able to detect the combinations of

assets and policies which have the greatest effect.
A MATTER OF PRIORITIES
Optimization can be used in a rather different way to suggest broad prior-

ities for investment. This is often illuminating, but it is also highly practical.
In many strategic analysis problems, whether they be of military or corpo-
rate planning, the problem proprietors do not know how much money will
be available to invest or , indeed, whether it is worth investing the maximum
which might be available; it is possible that the marginal benefit may not
justify the extra investment.
To illustrate this point, the combat model was optimized three more
times. In the first case, the budget was set to $8000.The optimal values were
then used as a new base case and the model was optimized again with a
budget of $8000, effectively making $16000 in all. Finally, the process was
repeated to give an effective investment of $24000. The models are on the
disk as OPT8 .COS, OPT16.COS and OPT24.COS. The idea is to show the
effects of increasing amounts of money becoming available , perhaps over
three successive budget years.
The optimal results are shown in Table 9.8, with the addition of the Total
Blue Losses, but how are we to interpret them?
The first thing to notice is a very clear priority to cut the strength of the
regular force by about 25% to release money for other purposes. The
reader should use the cost values given above to work out how much is now
spent on the other changes and should find that, when the budget
is $8000, the total money available from the budget and from reducing
the regular force is divided roughly equally between reducing BMDEL
and increasing RMCAP and IBLRES. When the budget is increased
by a further $8000, the money is spread between reducing BMDEL
and RMCAP still further, though there is a bias towards the former. Inter-
estingly, if there are no technical means of reducing RMCAP, we may
have identified a prior ity for Research and Development to find one.
This is an example of the way in which system dynamics experiment
and optimization sometimes triggers unexpected results in unanticipated
areas.
Finally, when a total of $24000 is available, the spread is between
BMCAP, RMCAP and IBLRES, with a heavy bias towards the reserve
forces. The move from $16000 to $24000 has involved allowing BMDEL to
increase because the increased numbers of cheap reserves only have an
effect if they can come into play. We are making the very unrealistic
assumption that investments can easily be reversed , but, if they cannot,
BMDEL would have to be excluded from the parameter plane in the 16000
+ 8000 optimization .
Before we leave the combat model, let us make two more comments,
which are generally applicable. First, we observe that the final result in the
16000 + 8000 case is rather better than when we simply optimized with
$24000. Specifically, the final blue strength is somewhat larger and the total
blue losses are rather smaller. The shape of the response surface is normally
so convoluted that more than one search strategy is usually worth trying.
The second is that there are no enormous differences between the results
from these different ways of spending $24000. The nature of response
surfaces makes it impossible to prove mathematically, but empirical evi-
dence from the use of optimization suggests that the shape is much more
like a corrugated roof than the peaks of the Alps. In other words, the
differences between peaks are not all that large that a sub-optimal solution
needs to be worried about too much, which also means that there is likely
to be a large number of solutions all of which are pretty similar in terms of
performance.
MAXIMUM PERFORMANCE OR MINIMUM COST?
In the previous section we maximized the performance which could be

achieved from the simple combat model for the expenditure of a given
amount of money. In many cases, however, one might wish to find the
minimum amount of money, or other resource , required to achieve a given
level of performance. A problem involving this appears in Appendix B.
MORE COMPLEX OBJECTIVE FUNCTIONS
In the DMC model, we used an objective function which recorded the

extent to which four variables failed to meet their targets ; for two of them ,
the target was obvious, and in the other cases we had to create a target , but
the underlying theme was trying to make discrepancies vanish. For the
combat model, we attempted to maximize a measure of Blue's ability to
survive; we sought to get as far away from 0 as possible, that is we tried to
increase a discrepancy. In short , the nature of the problem calls for some
thought in creating an appropriate objective function and, to illustrate that
style of thought, in this section we shall examine a case in which we need to
account for two aspects of performance, the first being the extent to which
an actual value is below its target and the second being the duration for
which the actual falls below the target. Practical instances of this might be
the extent to which hospital beds or fire engines are not available to meet
possible emergency demands.
Consider Fig. 9.13 (FIG9-13.COS), which shows a variable called
ACTUAL falling below TARGET at TIME=20 and continuing to fall until
it reaches 60% of the target. After that nadir, ACTUAL recovers until it
exceeds TARGET again at TIME=50. The fact that two targets , EIGHTY
and SIXTY, are plotted is to suggest that greater failures to meet the target
are more serious than lesser ones and should be penalized more heavily. To
see how this can be handled , let us consider the following equations, the
definitions appearing, as usual, after the equations.
note
note target values
note
200
/
/
//
/
/
'- '' ' .
125
/
""'.
100-1+;+-_ _---.::'"':- ----,''_---------
.
-----------~~-----~------------------------
"'". / .
------~------------
. . . . . .
II L TIME (WEEK! SII4lUTED TIME
' 1: TARGET (UNIT) TARGET VALUE FOR VARAISLE
1 -- ~~GHTY (UNIT) 801 OF TARGET VALUE
c:::.: Ac~0XL ~nl :~~ ~=OF VARAISLE
Fig. 9.13 Objective function for time and area .
a target.k=itarg
c itarg=lOO
a eighty.k=0.8*target.k
a sixty.k=0.6*target.k
These are some simple equations to set up a target and its 60% and 80%
values. The approach works perfectly well with dynamic variables.
note
note actual value
note
a actual.k = base - ramp(down,dtime) + ramp(up,uptime)
c base=120
c down=2
c dtime=lO
c up=6
c uptime=40
Again, this is a simple driver for ACTUAL, with constants defined to allow
other cases to be tested. The model on the disk tests ACTUAL under a
range of cases and writing a small, simple, model such as this to test an
objective function is an essential element of good practice .
note
note time storage
note
I below.k=below.j+dt*clip(l ,O,target.j,actual.j+eps.j)
n below=O
a eps.k = 0.00001 *actual.k
First, we store the time which ACTUAL spends below TARGET. The
CLIP function produces the value 1 whenever TARGET is greater than
or equal to ACTUAL + EPS. EPS is a very small number and is used to
ensure that the CLIP does not produce the value 1 when ACTUAL is equal
to TARGET, because that would be a circumstance which we did not
wish to penalize. As with all variables and functions in system dynamics
models, the CLIP is evaluated every DT, so, for every occasion on which
ACTUAL is below TARGET, DT is added to BELOW, thus recording
the required total time spent below target. In the base case run , BELOW
has the value 29.75, so the value of BELOW is DT too small, as the actual
value of BELOW should be 30. The reason is that level variables always
take 1 DT to be brought up to date, but the error is so small as to be
negligible. It could be reduced by making DT smaller, but DT should be
chosen to suit the delays in the system, not to make an objective function
more 'precise'.
To compute an index of the extent to which ACTUAL falls below TAR-
GET, consider the following rather complicated expression.
note
note area storage
note
I area 1.k= area1.j+ (dt/dt )*nclip((target.j - actual.j)**2,0,target .j,actual.j
x +eps.j ,actual.j,eighty.j)
n area1=0
COSMIC's NCLIP function takes the form:
v = NCLIP(A,B,C,D,E ,F)
and gives V the value A if C~D and E?;F. In this case it produces the result
(TARGET.J-ACfUAL.J)**2 if TARGET is greater than or equal to
ACTUAL+EPS and ACTUAL is greater than or equal to EIGHTY.
In other words it calculates a penalty if ACTUAL is within 80% of the
target. The penalty is multiplied by DT/DT when added to AREAL This
ensures that the dimensions of AREAl are [UNIT**2], rather than
[UNIT**2*WEEK). There is nothing to stop one taking the square root of
the NCLIP , to give AREAl dimensions of [UNIT]; it is largely a matter of
choice and personal style.
It may seem strange to say, as we have on several occasions, that some-
thing is a matter of choice. The reader may expect there to be more of
'science' and 'the right answer' in a book on modelling. That would, how-
ever, be misleading. Modelling is much more an art than a science and the
great skill is to model just enough of the problem to get a model which
is well suited to its purpose, though the model must also be soundly
constructed.
The line of thought used in writing the equation for AREAl is also
applied, in FIG9-13.COS, to AREA2 and AREA3, which calculate,
respectively, penalties for being below EIGHTY but above SIXTY, and
below SIXTY. The equations should be studied carefully. The difference
is that these shortcomings are penalized more heavily by being raised
to the fourth and sixth powers, respectively. Again, there is no magic
to this. If one felt that being below EIGHTY was no more serious than
being above it, there would be no need to use the higher power. If one
felt that it was very much more serious, one might use an even higher
power.
Using higher powers to penalize worse performance would not be valid if
the discrepancies were calculated as proportions rather than as absolute
values of the target. In such a case a suitable approach might be to use
(ACTUAL.JITARGET.J)**2, with the fourth or higher power used when
the ratio exceeds, say, 1.2 and 1.4. The reader should rewrite the above
equations to test this.
Clearly, care must be taken in formulating suitable objective functions,
but the point is that objective functions can be created to meet any desired
criteria for deficiencies in behaviour.
a total.k = areal.k + area2 .k+ area3.k
The three penalty terms are added together.
note
note the objective function
note
a objfun.k = (belwt *below.k/scale1 + areawt*total.k/scale2)/(belwt
x +areawt)
The final objective function is now the sum of the area and duration penal-
ties, suitably weighted and scaled to make each equal to 100 in the base
case. The final division by BELWT plus AREAWT makes OBJFUN also
equal to 100.
The reader should study the model on the disk, noting that all variables
are printed out and that a series of test cases are used to verify that these
equations work properly.
note
note definitions
note
d actual=(unit) actual value of variable
d areal = (arbunit) penalty component if actual is less than target but
x more then 80% of target
d area2=(arbunit) penalty component if actual is less than 80% but
x more then 60% of target
d area3=(arbunit) penalty component if actual is less than 60% of
x target
d areawt=(l) weighting for relative importance of being below target
d base=(unit) base level for actual value
d below=(week) time spent below target value
d belwt=(l) weighting for relative importance of time spent below
x target value
d down=(unit/week) downward slope for actual value
d dt=(week) solution value
d dtime=(week) start time for slope down
d eighty=(unit) 80% of target value
d eps=(unit) small number to ensure correct comparisons between
x target and actual
d itarg=(unit) target to be achieved
d length=(week) simulated duration
d objfun=(arbunit) value of objective function to be minimized
d scale1= (1) scaling factor for time spent below to ensure objective
x function = 100 for base case
d scale2=(1) scaling factor for amount actual is below target to ensure

x objective function = 100 for base case
d sixty=(unit) 60% of target
d target=(unit) target value for variable
d time=(week) simulated time
d total=(arbunit) total of area components in objective function
d up=(unit/week) upward slope of actual value
d uptime=(week) time at which upward slope starts
FACTORS IN CHOOSING AN OBJECTIVE FUNCTION
The enormously wide range of applications of system dynamics makes it

hard to give much prescriptive guidance on the selection of a suitable
objective function other than
• to think carefully about the managerial problem
• to test the proposed equations to make sure that they work correctly
• not to be limited by the software
The need for thought is illustrated by the DMC model. A naive analyst
might be tempted to minimize Order Backlog, BLOG, but that would be
to undermine the whole basis on which DMC deals with its customers.
Similarly, minimizing Raw Material Stock seems like a good, cost-reducing,
idea. Unfortunately, the minimum stock is zero, and, with that level
of stock, production would have to stop . In both those instances, a little
more thought makes it clear that keeping Backlog and Raw Material
Stocks close to reasonable levels is what is required by the nature of the
managerial problem, and it is that nature which the objective function must
reflect.
Similarly, a little thought reveals why minimizing cost is rarely suitable as
an objective function . In the DMC problem, one might well be able to
calculate the costs of varying the Production Rate or of holding Raw Mate -
rial Stock, and minimizing those costs would be well within the mainstream
of management science thinking. If, however, one then ignored the impor-
tance of keeping Backlog close to its target, one would have missed out vital
aspects of the managerial problem and it would be verging on fantasy to try
to ascribe costs to the difference between Desired Backlog and Backlog.
The essence of the matter is that the objective function should reflect what
the problem proprietors think are the important attributes of system per-
formance and, in management, cost is usually only part of their thinking.
Similar considerations would apply to the simplistic maximization of profit,
not the least of the difficulties being the difference between the short and
long terms.
Minimizing cost might, however, be appropriate in constrained optimiza-
tion. In the combat model, one might seek to minimize the cost of achieving
a given level of Blue victory rather than trying to find the maximum Blue
victory which can be achieved for a given amount of money (See Appendix
B, problem 15).
Similar ideas apply to problems of corporate investment.
The importance of writing a simple model to test the workings of a
proposed objective function with very simple inputs was demonstrated in
the previous section. This is usually a very good strategy because the objec-
tive function is normally only formulated after the model has been devel-
oped, revised and experimented with. In practical analyses, one often has
rather a large model, perhaps hundreds or even thousands of lines of
equations and definitions, by that time. Attempting to use that as the basis
for testing objective function, equations which, by their nature, ought to
reflect some subtle ideas and, perhaps, embody some complex equations, is,
at best, cumbersome and, at worst, prone to error. Having the patience and
self-discipline to pause to develop a small model to test the objective
function is good practice.
Finally, it is important to allow oneself full rein in writing equations
which adequately capture the various factors in the objective function .
There are no limitations on what can be modelled, so do not merely select
some simple variable and simplistically optimize it.
For instance, had we chosen, in the DMC problem , to minimize the
difference between BLOG and DBLOG and written a component of the
objective function as:
a penl.k=dblog.k-blog.k
we would have run into the problem that the minimum of PENI is when it
is as negative as possible, which will happen when BLOG is very much
larger than DBLOG, an absurd idea for this company. Using the absolute
value:
a pen l.k = abs(dblog.k- blog.k)
would be no better because the software calculates the objective function
only at the end of each iteration. That would mean minimizing PENl (and
the sum of the other penalties) only when TIME = LENGTH, and paying no
attention at all to the dynamic behaviour over the length of the run; equally
absurd.
THE SIGNIFICANCE OF THE OBJECTIVE FUNCTION
Before moving on to some final technical points on optimization, this is an

appropriate stage to understand the true nature of objective functions.
Towards the end of Chapter 4 we discussed the significance of DT and
emphasized that DT is, in a sense, a figment of the modeller's imagination
which is required only for the purpose of getting the model to run on a
computer. It has nothing do to do with the real system, though its numerical
value has to be chosen correctly.
Somewhat similar considerations apply to objective functions, in that
they are an act of the analyst's imagination and are required in order to get
the optimization software to run. Unlike DT, the objective function does
have something to do with the real system, because it penalizes undesirable
aspects of performance and rewards beneficial components.
It is, however , in the highest degree unlikely that, even with all the
thought and care a skilful analyst bestows on formulating an objective
function , it will completely capture the nuances and subtleties in the minds
of the system's proprietors about what does and does not constitute good
performance. In short, the objective function should not be accorded a
status it does not deserve as the final arbiter of policy and structural options;
that role belongs to human judgement of the dynamic behaviour those
'optimal' policies and structures produce. Put another way, the objective
function should not be taken too seriously. Indeed, one should be prepared
to abandon an objective function and create another if it triggers thinking
about different, and better, ways of gauging performance. An objective
function which does that will have served its purpose very well indeed
because the underlying value of optimization is to stimulate deeper thinking
about the problem than might otherwise have occurred.
NON-LINEAR OPTIMIZATION
Non-linear functions are common in system dynamics models and , as ex-

plained in Fig. 5.4, are normally represented as TABLE functions, or one of
the variations thereof. As we saw there, a TABLE function takes a form
such as:
T TFSRF=O/O.024/0.12/0.8/0.98/l.0
That particular table represents the effects of food supply on the longevity
of foxes, but a similar construction is used in the pharmaceutical company 's
model in Appendix B to represent management's preference for allocating
expenditure between two competing demands.
A TABLE statement is clearly a list of numbers written in sequence, in
other words it is a special form of a constant. As such, it can perfectly well
appear in the parameter plane if the optimization software is powerful
enough to accept it. In order to provide the ranges within which the table is
to be allowed to lie, one simply has to specify a pair of numbers for the
upper and lower limits of each entry in the table . The COSMOS optimi-
zation software will then treat those as parameter limits and, in effect,
design the shape of the curve which gives the optimal performance (in
conjunction with any other ordinary parameters or tables which are in the
parameter plane). Since this book is concerned with the principles of system
Fitting models to data 295
dynamics rather than with any particular software support tool, we shall not
discuss this any further , though the topic is explained in detail in the
COSMOS User Manual.
FITTING MODELS TO DATA
In some cases, data are available about the historical dynamics of a prob-
lem, and some of the parameter values are uncertain. For instance, in the
DMC problem one might have data about New Orders over a two-year
period and the corr esponding records for Production Rate over the same
period. In such a case, one could set up two tables. That for NO would look
like:
a no.kl=tabhl(tno,time.k,0,24,1)
t tno=120,145,163 . ..
Incidentally, the same method could be used to set up the Blue target,
BTARG, mentioned above.
A similar table could be used for Recorded Production, RECPR. An
objective function could then be:
I pdiff.k= pdiff.j +dt *(pr.jk- recpr.j) **2
to minimize the difference between recorded production, RECPR, and
simulated production, PR , when parameters such as TAPR, TABL etc. are
allowed to enter the parameter plane . Care would have to be taken to get
the initial backlog and raw material stocks correctly set to data values and
to load up the delay in the raw material pipeline correctly, in that case
by disaggregating the internal levels of DELAY3, as was discussed in
Chapter 5.
The difficulty with this approach is that there may be more than one
series of data; perhaps Raw Material Order Rate records also exist. Does
one give more weight to matching the production data or the raw materials
records? How does one know that the data are correct? Some of the
parameters might be known with some confidence, in which case only those
which are uncertain should lie within the parameter plane .
In short, there is nothing wrong with fitting models to historical data,
though the problem is rarely as straightforward as might appear; the
author's experience is that there have very rarely been occasions when one
had enough confidence in the data to fit anything to them.
Above all, one must avoid thinking that one has proved , from historical
data, that the model is 'valid', let alone 'right'. The reader should pause to
review the comments made in Chapter 6 about the validation of models.
Fitting a model to data does give some added confidence , but it is only part
of a much deeper problem. One must also remember that system dynamics
models are intended for policy design, and one is less concern ed about what
the parameter values now are than with what they should be to get good
performance from the system. Most managed systems behave quite poorly,
and there may be little point in carefully using optimization to discover the
parameter values which give bad performance. It seems rather more con-
structive to use optimization to design policies and structures which give
good performance.
SUMMARY
In this chapter we have sought to expose the reader to some of the range of
powerful ideas in system dynamics optimization. We have discussed the
vital role played by experimentation to find the range of options which
might enter the optimization search and we have considered the formu-
lation of objective functions. The ideas were illustrated by two cases of the
different types of optimization. We laid great emphasis on simulation by
repeated optimization, showing that optimizing the same problem in differ-
ent ways could yield quite different results.
It is, however, absolutely essential to be clear in one's mind about what is
being done. The aim is not the optimization of a model, it is the design of
management policies and structures. Computations to search a parameter
plane are really attempts to search the 'management plane '.
A major point was the idea that optimization is very powerful but, like all
powerful tools, needs to be treated with care. Simply optimizing the obvious
parameters without much thought is likely to lead to unimpressive reSUlts,
no better than can be achieved by intelligent experiment. Optimization
after experiment usually, however, leads to considerable further gains in
performance.
In short , optimization is probably the most powerful development in
system dynamics since Forrester's original stroke of brilliance in inventing
the field. It is, however, neither a panacea nor a substitute for thought, and
the objective function should never be allowed to be the final arbiter of
system design; that role belongs to the judgement of analysts and managers,
and more to the latter than the former.
CHAPTER 10
Advanced modelling
This Chapter deals with a number of special modelling techniques which

will be required by the advanced student or the serious practitioner. For
example , it explains how to model changes in a variable, including discrete
changes. Achieving controls within limits are described, as is modelling of
market share . Techniques for the modelling of forecasting are explained. A
section is devoted to equations for financial accounts and balance sheets .
The important subject of the correct use of multipliers is studied in detail.
Several other techniques and examples are included. Additional advanced
modelling techniques and cases are described in Equations for Systems,
mentioned in Appendix C.
Such material logically falls into this part of the book but, to avoid
distraction, the text of the chapter is included on the disk provided with the
book . The models referred to are also on the disk, numbered as mentioned
in the Chapter, as are the diagrams (in CorelDRAW format). For readers
who do not have access to that excellent graphics package, the diagrams
follow. In themselves, they are meaningless without the text, though they
will give a flavour of what the text covers.
298 Advanced modelling
Upper Limit UUM
I
I
I
I
I
I
I
I
I
Lower Limit LUM
----------,-------------
I
I
I
I
I
I
time
Fig. 10.1 The dynamics of production - case 2.
YearI Year2
• •
B1
\
PPSR for Yeor 2
~ f------------- ---
PPSR for Year 1
- - - - - - - - - - - - ---
A2
I
I B2
A1
\
• I
I
\
~
r--!-
I
FFT DMT DEl FFT DMT
3 10 12 15 22 TIME In Months
Fig. 10.2 Freezing of forecasts.

Advanced modelling 299
Assets Liabilities
( DAR-DRR) _ - - - - - - ~e
- - - - NSlR- - - - -- - - -+~8
DIV- - __
.... ,
/ TAXR ..... ,
+..- '
.,
TAX _ _ _ "
~ '\ --- \+
\
,\ --~~RTP
'\
\
\
+1'
I
\
\
\ I \
ORR-IPR).... \ I \
" \ , I
'\ \ / ~
....
~ +\ 1 /
.... ~ , /
-:Tf:-/
(
CCf~
- / I
RC / I
+ / I
// /
SR- - - / /
/ .:
+,! _/
~-------------:f?
--- MBPR~
Fig. 10.3 An influence diagram for a set of accounts.
FullDesign Capacity
10
Threshold
.,,
\
\
Read iness Reac hes
Threshald and Plant
Is Regarded as
Having Reac hed
Readiness FullDesign Capacity
Builds up - - --.
~
/ Time
/
/
/
Plant Ordered
Fig. 10.4 Plant 'readi ness' and output.
Table 10.1 Components of a balance sheet
TA Total Asse ts TL Total Liab ilities
CASH Cash SHC Shareholders'

Capital
+ +
RC Receivable RTP Retain ed Profit
+ +
MVI Monetary DEBT Total Debt
Value of
Inventory
+ +
FA Fixed Assets CL Current
Liabilities
6000
11
I
I
5000 I
4000
3000
2000
1000
I
I
lIO .
L
CPBF
--BFPR
TIME IDAY) SIMULATED TIME
(TON) CUIfJUoTIVE PAOOUCTIDH FROM BLAST F~CE
(TDH/DAYI PAOOUCTIDH FROM BUST FURNACE
15 .,
I
12 ·sJ
I
lo·i, 35'
30.1'
,
7 . sJ
5·i, 25 .
r----' r----\ ,-----\ ,..-----,
2.&1 20 . I
I I
\ : I
:
I
'
I I
I ,
I
I ____ II " J L J I I\ I
~ \1
O.~ 15 .
I
-2 .&1
I 10 .
-5 .~
I
5.
-7 .sj
I
-10 .J . -1,"6-. -J,.,,---~-~r--',..---l:r...------1~0
0 'if-.- --lx-..,....-\orrr-J....,,- 20. 28. ~2. ~6 . .
TIME IDAY' SIMULATED TIME
!--------. INVBLF
L HTMBF IDAY) OU...Y VARIABLE FOR TIME RELINE STARTED
(DAY) INTERVAL FOR SAMPLING THE BLAST FURNACE
TRI AL OF MODEL
Fig. 10.5 Test of blast furnace equations.

A-- __
.........
<,
<,
\+
+ •
B-------. X
~
/+
/'
C----- .-
etc etc
A Row Relationships B OtherRelationships
Fig. 10.6 A simple influence diagram of multiple relation ships.
1.5
g
Q
0
o
CD
c
~ A - -Curve
---C
~ 1.0
E Linefrom
g Equation 82
\
.gj \
a. \ I
~ I
~
c
0.5
'" I
I
I
~::J / Curve B
s
a. /
/
/
/
--
/
0.5 1.0 1.5

Ratio of SPR to PPR
Fig. 10.7 Produ ction multiplier from working capital.

1.0 .. ------..- - - - --
a '--__~-------__:;::_;::.
0.5 1.0 1.5 2.0
SRAT Number of Shops Relaflve to Soturaflon Level
A The Satulatlon Muff/pileI
1,0
OL..-~=--- _
a STPS Stock per Shop
1.0
B The stock Muff/pileI
~ ~I.. __ TooManyCooks
./ ....uttIe Effect
fQ I
I
I
I
I
1.0
ARERActual to Required Employees per Shop
C The Employees Muff/pileI

Fig. 10.8 Three cases of multipliers.
+ =\+ //
"",..-- ........
r "'-"--. ~~'fng
/ +"'Alrcratl Loss
/ MIssIon
\ I
/
+ Operational
_____ A1rcratlln ~
" C'-
Area --
'\I
Losses
Rate of
Entering +,. DutaIIon
~
Ope<atIonal Rate of _oI00nger
Area Leaving _ - - - Pe/fod
Oln&OONO o~~~
C
Alrcr~:""'h'" -- <, <, RETURN AlrJo:
, In Retum
+- Stage of
T M~~
/
I +~-
Alrcratl
_ Aircraft Available Retum
"~~~)Rate
Maintenance
~ _
Aircraftln
Maintenance +
Fig. 10.9 Influence diagram for detailed approach to aircraft sortie problem .
100
90
80
70
60
50
40
""
30
II
II
20 I I 1\
I \ / \ ......., _
10
~ -------------
L TIME (HOUR) SINlL'TED TIME
L_ ~~~~HL 1~~=H ~~~~~ ~~A~~~I~~a;~TlONS

TRIAL OF MODEL
Fig. 10.10 Aircraft sortie model.

100
90
eo
70
60
50
40
30
20
10
c.
/J""",,,,,,,,,V',,V',,V',,V''''''.IV'.I''''''''A''\./,/\/J''\ I \ i\ r. r. I' t,
o . ~o . 20 . 30. ~o . 50. Bo. to.' ~o~ • bo ." \00 .
L TI ME (HOUR) SIMULATED TIME
ACAVAIL lAIRl:AAFT) AIRCRAFT AVAILABLE FOR OPERATIONS

-- OPAREA AIRCRAFT) AIRCRAFT IN OPERATIONAl. AREA
A - GROUPS OF 20
100
90
eo
70
60
50
40
30
20
10
L' . TIME (HOUR) SIMULATED TIME
ACAVAIL (AIRCRAFT) AIRCRAFT AVAILABLE FOR OPERATIONS

-- OPAREA (AIRCRAFT) AIRCRAFT IN OPERATIONAL AREA
B - GROUPS OF 50
Fig. 10.11 Pulsed aircraft sortie model.

CHAPTER 11
The applications of system

dynamics
INTRODUCTION
This chapter indicates the enormous range of applicability of system

dynamics to help you to imagine how a problem might be analysed from
a system dynamics viewpoint. We cannot hope to describe every single
application or mention every worthwhile example; the latest bibliography
of published work on system dynamics lists over 3000 references'.
We start with a brief history of system dynamics. There follow a few
case studies of system dynamics analysis. Finally, the defence applications
of system dynamics, which have been particularly successful, are surveyed.
The disk at the end of the book contains listings of the models used as
cases.
These models should be run to produce the graphs which are mentioned,
but not shown, in the text (the diagrams are numbered to allow for this).
This is vital practice in working with models and the student should not
neglect it. In some cases, the model will offer interesting possibilities for
optimization if the software being used supports that facility.
SOME SYSTEM DYNAMICS HISTORY
System dynamics originated at the Massachusetts Institute of Technology in

the late 1950s thanks to Professor Jay W. Forrester.
Forrester's first book was entitled Industrial Dynamics (Forrester, 1961),
though the wider term of system dynamics was soon coined. Work of such
novelty led to the development of a number of models of industrial prob -
lems (see e.g. Roberts (1978)). These models are, however, completely
closed unlike the DMC model, discussed in earlier chapters, in which
the behaviour arises from the interplay between endogenous factors
1 Available from The System Dynamics Society, 40 Bedford Road , Lincoln, Massachusetts
01773, USA. A search of that bibliography will produce numerous reference s in addition to
those described here.
Some system dynamics history 307
and exogenous driving forces (for some models of this type, see Coyle
(1977)).
System dynamics, despite its origin in business management, was clearly
applicable to social problems, and this has proved to be a fertile area of
application.
In the late 1960s, American society was perceived to be in crisis and
Forrester took up the challenge by presenting a model of the dynamics of a
city.
Urban modelling was a strong theme in system dynamics for a few years.
Refinements of the urban model were developed and applied to particular
cities. Further publications were produced (see for example Alfeld (1975),
but the application of system dynamics to urban problems seems to have
fallen out of favour.
Unfortunately, that syndrome was repeated with the publication of
models of the future of the world system in the 1970s. Supported by
a group of concerned politicians and business leaders, The Club of
Rome, Forrester produced World Dynamics (Forrester, 1971), the model
being created during the flight from Boston to Rome! This was ac-
complished without the aid of laptop computers, perhaps suggesting that
the real skill in model development is thought, not dexterity with software.
The model extends to no more than four pages, including the definitions
and dimensions of the variables. A second phase of the work was the
WORLD2 model of Meadows et al. (1972). A full description of subsequent
work on WORLD3 appears in Meadows et aI., and their influence diagram
appears in Fig. 11.1 to indicate some of the characteristics of the world
models.
The diagram reflects only what its authors saw as the most significant
aspects of the model, so it is drawn without all the detailed conventions of
influence diagrams ' . To show the most striking assumptions, one being that
Industrial Output per Capita decreases Fertility, whereas, on the other
hand, total Pollution (not per capita) increases Mortality. Is this an incon-
sistency?
We do not aim to answer such questions, merely to show the reader how
the careful study of a diagram can raise them. If one's assessment is that the
diagram is correct, confidence in the model has been increased. If it is that
the model is not correct, the model does not become useless. Either one has
identified an area in which it can be improved or one has identified a
limitation which can be taken into account when the model's policy impli-
cations are evaluated. The reader should study the subsequent case studies
from this perspective. No model is 'right' in any absolute sense: the judge-
ment is whether it is right enough to be useful.
The widespread emphasis in the 1970s on environmental problems natu-
rally generated much further system dynamics work. An interesting exam-
2 Users of causal loop diagrams do not adopt the dotted and solid line conventions .
~P~aIi~
on~hsper
I. ) Yea, , +
Fi~~~ + t~+
-, Education. Heallh
+
FamilyPlc.-ming SelVices Food ~
Cultiva ted
Industrial + ~ + land
Agricuttural
Investment +
Oulpulper _
C",".
~ \ _~
7-
+ ~ Capita
__
D;:;~a~'fad +.. + Ca pit
~____
<: \ al + .
SelVice / +
+TaI ~Pollution
output
- + +
Nonrenewable
Resources
<:>. Industrial ~+
c a plt, _ n .
(
. of Industriol
.> ~e~ed~on
+
C~
'~-_m~ ~_ ......
..
+ I + of Ca pital
Investment
Rate
Fig.11.1 Influence (causal loop) diagram of structure of WORLD3 (afte r Meadows et al. (1974)).
pIe is a model of the transition from fossil fuels to sustainable energy

sources (Naill, 1977).
In short, a number of themes had emerged from these early models. The
first was that the concept of using the principles of control theory to analyse
managed systems via a simulation tool had proved to be very practical. The
second was that the fields of application were limited only by the wit and
imagination of the analyst.
The dynamic behaviour of a wide range of systems could be modelled
without much difficulty, and the policy implications were often striking and
novel. Unfortunately, they were often also bitterly controversial. The
source of controversy was that most people do not think in dynamic terms,
and having someone else analyse 'their' problem from a completely novel
standpoint was not always comfortable, especially so if apparent errors or
misinterpretations of established theory could be pointed out',
Current work is covered in a refereed journal, System Dynamics Review
and the Proceedings of the annual conferences of the System Dynamics
Society'.
We now turn to the more detailed examination of some system dynamics
case studies.
THE NEW ZEALAND FORESTRY INDUSTRYs
Introduction
The economy of New Zealand is significantly dependent on the export of
timber in various forms. Large areas had been planted during the 1960s and
1970s which meant that there would be a large crop during the 1980s and
1990s. Continued planting would increase that crop. However, there were
risk factors, such as intensified competition in world markets. Conse-
quently, in the early 1980s there was a need to design planting and cutting
policies to cope with an uncertain future.
The suitability of system dynamics

How do we know that this is a problem for which system dynamics is more
suitable than, say, linear programming, which has been applied to the
optimization of land use? There are a number of factors:
3 J. H. Plumb, Boxer (1965), writing of the need to make integrated sense of history, remarks
that 'The majority of historians .. . have spent their leisure tearing to shreds the scholarship of
anyone foolish enough to give the story of mankind a meaning and a purpose . Writers as
diverse as Wells and Toynbee have been butchered with consummate skill. .. . the errors of
interpretation have counted for everything ; intention nothing'. One feels that could easily be
rephrased to apply to the early workers who attempted to apply a new paradigm to the
management of social and economic systems.
4 For information on membership of the System Dynamics Society or subscriptions to System
Dynamics Review, contact John Wiley & Sons Ltd, Baffins Lane , Chichester, West Sussex
P019 JUD, UK.
5 For a fuller discussion see Cavana and Coyle (1984).
310 The applications of system dynamics
• There are exogenous shocks, such as changes in demand.

• There are long delays, trees taking about 30 years to grow to harvest size.
• There are dear feedbacks: the more trees are planted, the fewer will need
to be planted to meet some anticipated future demand.
• There are significant policy issues for the long term, as opposed to short-
term 'tactical' decisions.
In other words, a problem as vast as a nation's forestry policy can be

interpreted in terms of the information/action/consequences paradigm
shown in Fig. 1.1. In addition, the concerns of the nation's policy-makers are
the four upper topics on the right-hand side of Fig. 1.5.
The main point is that an analyst should always justify the use of a
methodology. To do so requires familiarity with other methodologies, and
doing good system dynamics requires at least a working knowledge of other
management science approaches.
Some simplifying assumptions
Any model is a simplification of reality, and the art is to ensure that there is
nothing in the model which is not also in the real world and that the
significant parts of the real world are in the model. Some modellers try to
put as much as possible of the real world into their model. This leads
to models which are intractable and difficult to interpret. The key step is to
make some assumptions which are sensible for the problem and then to
make sure that those assumptions are faithfully adhered to.
For this model there are two main assumptions.
The first is that trees can be grouped into four categories:
• Young forests less than 10 years old, which are non-productive and
require attention for pruning and thinning.
• Immature forests (10-25 years old) which could be felled. However, New
Zealand's plantation forests have usually been managed to produce logs
from trees at least 25 years old, so the immature forests are treated as
non-productive. Altering the model to vary this assumption would make
an interesting project.
• Mature forests (25-40 years old) which contain wood that is at the
desirable age for clearfelling.
• Overmature forests (more than 40 years old) containing slow-growing
trees beyond the desirable age for clearfelling.
As a rule of thumb, the LENGTH for a model should be about two or three
times the longest delay. If trees are in the system for up to 40 years,
LENGTH will have to be about 100 years. This matches the reality that
forest management is a very long-term business.
The second assumption refers to the price of wood. If demand for wood
and the rate of felling are identical, the price is $10 per cubic metre. If the
output of wood is less than the demand the increase in price will be small.
If the output is greater than demand, the price will fall more sharply.
The physics of the system

We have emphasized the need to identify physical flows correctly, and the
first assumption identifies the first physical flow of trees on a given piece of
land moving from one age category to the next. This is, of course, exactly
the same as for manpower planning models.
However, older trees contain more wood than younger ones, so there is
a matching flow of volumes of wood.
Finally, what is sold is volume, which will determine the area to be
clearfelled dependent on the age of the trees when they are felled. Thus, the
third flow is land being planted, cleared, sold or purchased.
The influence diagram

You should now develop your own influence diagram. Having parsed,
one might use the entity/state method for a clearer picture of the physics.
The list extension method should also be used to think about control
policies, and the two fragments should then be merged into a candidate
diagram.
Figure 11.2 shows the resulting diagram at about level 3. The first and last
two categories of trees have been grouped together to show the ageing of
trees. This Area Transfer Rate creates the Volume Transfer Rate to feed
the volume of wood which is available for cuttings, The volume is also fed
by the Volume Increment, which is the continued growth of trees in the
Mature and Overmature classes.
The need to ensure a balanced age structure in the forests implies
the Ideal Clearfelling Rate (in hectares per year) and, through Average
Volume per Hectare, gives the Ideal Cutting Rate (in cubic metres per
year). However, Demand and Price also govern the actual Volume Cutting
Rate.
The Planting Rate must replace felled areas and meet expected future
demand. That leads to the land flow in the lower half of the diagram.
The reader should add a balance sheet, using the techniques discussed in
Chapter 10.
The model equations

WOOD.COS should now be carefully studied, with the aid of a printout. It
would also be useful, if the available software provides such features, to
6The volume of young and immatur e trees need not be considered , as they are deemed not to
be capable of being harvested.
-- -- _- - - -1~I- - - - .... .... ,

'.
/~~ I ~+
'" I _-- PrIce
",'" I ~ - otWood
[,=
'"
11+ Idea
+ Cuffing - - - .. Recommended
_Y Rate
+ +t.,+
Cut
I
,
/ t",Wood '" - It + \ I
/ ",'" \ , I
/ / \ ", I
I / \ +
+ .. \ Vol""" \ + Volume of .... --~
,Tr~tet' ~~~
"FJ"'e<f
Desired .. t4- Cuffing
S~~
/Areoof
',:Jf ,....
e
+
<,
'
Rate <,
"
/ Foreol1 Total / . . . . .... .... , <, \
Stocked I ........ I d e a l " <,
I _- - Area ot I Oeorfeiling " " \
I / F",eol1 .. Rate ' \
I .. \ + I / ~+, \
I /
+1
,,\ "*'I
/ \ Volune '
Increment " ..
1.+ 1
\, I I" \ o.lay \.+ <, AVIlfage I
j. I"~ \ I .... - V~ _
+ll - I x> ' \ I Hectcre '\ J
Gap
I
I
I \',
I'....
\ I
\
\ I
-t
",/ /
I
I
I
I I "........ I -It /
'Area --~~--_......
\
\
Arec or
Young and •
lnYnature _
Transfer
rate
~ Overmature •
+ Trees _
Area
Oeartelling
__ Rate
.-+/
\\" <-l<"'t'~,,-j \~ ,:~>/

, Av_hg /
c_ ,+ rime - +y
r.m. - - ~AOntlng - ~C~~~
/
'" Rot~+
, /
Rote
AV~ I ....... , ---.-/

1JfM - _ ++ _ land
.... ~ Ave<0g8
~
Rate _ - -
-------- AVailable~ ~ Planting -
~
1 -i +
t _ - - - .. Gop
+t / - + \ ~PIJr~e ~
- ....~:~~ \ /" Rate r~_
''''Acn~ng +.
\
I
Net La'ld
Puchose Rote
/ ~ :-/ '
- Land Available +
~ Purchceng
I
/
'" ",Y \ t",Foreofty /
~"''' '"
------------
I'imt....... ,. ",
Fig. 11.2 Simplified influence diagram for the forestry model.
produce a printed map of the model's interactions, a documented set of

equations (see Chapter 5, p. 136) and a dimensional analysis.
The model has several of the characteristics of good practice, notably
comments within the equations, full definitions of all the variables, and
extensive print statements, providing evidence that the model has been
carefully debugged. Several plot statements have been suppressed, though
again they suggest thorough debugging, only those for the main aspects of
performance remaining.
Numerical data are from New Zealand Government statistics. The initial
value for the Average Planting Rate reflects the previous policy of planting
as much as possible. When the model starts , a more rational planting
policy is introduced, with consequences which we shall see in the model's
output.
The equations for profit and cash flow should be studied, as should
the sector headed 'performance indices'. The numerical values for con-
stants such as CWRFV have been found by running the model for the base
case and finding, in this case, the value of Cumulative Wood Removed,
CWR, at the end of the run. This is then used as a base value to assess
the effects of other policies. No amount of further explanation by the
author will help the reader to develop the practical skill of grasping what a
modeller has done; you will have to work these equations out for yourself.
The reader may wish to modify these performance indices, or to create
others .
Another indicator of good modelling practice is the long list of sensitivity
test experiments, suppressed by NOTE statements. The reader should reac-
tivate some of these and see what they do.
Running the model
The reader should now run the model' .

The base case, Fig. 11.3A, assumes medium demand growth at a constant
rate of 0.3 million cubic metres per year. In these conditions, supply
matches demand quite well in the long term. However, the heavy planting
of the 1960sand 1970scomes into effect for 20 years from about 1995,when
the available cut from the forests, ICUT , is much larger than demand. The
pricing mechanism allows wood sales to increase somewhat , ACUT, but
wood prices are depressed for about 25 years.
The previous planting history means that APLANT is much higher than
PLANT, but eventually the two settle down. Land is sold during the 1980s,
but land purchases stabilize towards the end of the run.
With lower demand, Fig. 11.3B, the lower prices last for about 50 years,
and the sale of land continues for a long period. Even with high demand,
Fig. 11.3C, there is still a period of price depression, but it is relatively short-
lived. A problem looms towards the end of the run when land purchases
cease because there is no more land left to purchase . Increase LENGTH to
150 years and explain what happens .
These runs are not, of course, forecasts; they are assessments of the
future consequences of varying events. All show the serious implications of
7 The reruns can, of course, be done inter actively on a PC, but they have been left in the model
for instructional purposes . Instructors on graduat e or undergraduate courses will be able to
develop some inter esting individual student projects or class assignments using this model, or
the others which appear elsewhere in the book .
the past policy of heavy planting. Even if high demand can be stimulated, a
period of depressed prices is inevitable and the distant consequences of
high demand are pretty unpalatable. Whatever happens, there are likely to
be heavy costs from the previous policy of intensive planting which may
well have serious knock-on consequences for the rest of the New Zealand
economy.
The model forecasts demand for wood 30 years hence", The three cases
examined so far assume that demand is forecast perfectly (study the model
at about line 23 to see how forecasting error can be modelled). Figure 11.3D
shows the effects of forecasting errors, the pattern tested being pessimism
when demand exceeds the actual felling rate, and vice versa. With the
medium demand scenario, this introduces cycles of behaviour. For instance,
the over-supply of wood during the 1990s causes managers to take a pessi-
mistic view of the future and the planting rate drops. The eventual short-
ages of wood cause planting rate to rise, and it is not until towards the end
of the run that the control policies restore stability.
Numerous policy experiments can be tested. Such as a rotation age of
37.5 years, and an alternative planting policy.
The performance indices mentioned earlier should be printed out and
tabulated against the various runs as a way of assessing other effects of
policy as well as what can be seen from the dynamic behaviour. (See also
Coyle, 1981a.)
Summary
This model illustrates the process of turning a policy problem into a policy
model and several aspects of modelling method. The reader should study
the equations, and run the model, carefully studying the output'.
What other assumptions have been made? What modelling
simplifications have been made? Are there any other policy options?
Draw an influence diagram at Levell to reflect the insights derived after
the model has been carefully studied. The reader who has access to opti-
mization software will be in a position to explore this model very deeply, on
the lines shown in Chapters 8 and 9.
Technically, this model is very classical system dynamics. It is essentially
continuous in that land, tree age groups and timber volume do not change
in sharp increments. We now turn to a policy problem for which those
assumptions would not be reasonable.
8 A map (p. 137) of the model will help immensely in tracing out these factors.
9 This remark may seem superfluous. The reader will just have to take the author's word for
it that he has seen many cases in which the modeller in question had done no more than glance
at a few graphs on the screen and had completely failed to see errors which became obvious
from even a cursory examination of the output and which undermined any credibility the
model might have had.
DEVELOPING ARGENTINA'S ELECTRICITY CAPACITYIO
Introduction
The purpose of system dynamics is the study of the evolution of a socio-
economic system through time and this often requires the modelling
of capacity expansion. National energy supply policies are clearly such a
problem (see also Naill (1977».
Argentina's energy policy in the late 1970s aimed to increase the use
of hydroelectric generation. This programme was to be implemented
over a long period, in the face of unpredictable changes in the price of
oil and geared to meeting an increasing demand for electricity (for
an interesting discussion of future prospects in Argentina, see Macrae
(1994».
Earlier system dynamics work on energy problems has often been based
on models in which capacity can be added in continuous quantities. Hydro-
electricity is, however , a case in which the generating capacity available
from places where dams and power plants can be built varies enormously
from one site to another.
In Argentina there are about 26 places at which hydroelectric power
stations could be built , but their potential output varies from a few hundred
to nearl y 20 000 GWh (GWh ) per year. The costs of development vary at
least as much and the time required to build a station will also range
between roughly 9 and about 12 years" . However, the output from the
stations will be a continuous flow of electricity which will produce continu-
ous flows of cash.
For credibilit y, each project must be treated individually, which requires
the modelling of wheth er or not to order a parti cular project at a given time
together with the continuous processes of the output of electricity and the
flows of cash as revenue and expenditure during construction. Such com-
plexity has often been assumed to be impossible within the framework of
system dynamics software packages, but work on modelling coal production
has suggested that the difficulties were not as severe as had been assumed
(see Equations for Systems, described in Appendix C).
The reader should now think through the justification for using system
dynamics for this problem.
10 For a detailed discussion see Coyle and Rego (1982).

IIThe four river basins are the Plata, on the international border with Brazil, Uruguay and
Paraguay, the Patagonian basin and two basins in the Ande s, Cuyo and Neuqu en . The Plata
projects are large, with high construction costs and low unit output costs. Construction delays
are about 12 years. The Cuyo system offers small projects, with relat ively low construction
costs but higher subsequent operating costs and constructi on delays averaging about 9 years.
The main difference between the Patagonia and Neuquen basins is the likely constructi on
delays, abo ut 12 years for Patagonia, which is about 2000km from the main electricity mark ets,
and 9 years for Neuqu en.
Capacity expansion policies
Figure 11.4 shows some possible policies. If capacity can be added in effec-
tively continuous quantities, the idealized case in Fig. 11.4(a), projects will
already be in the pipeline and demand will be forecast for the construction
delay into the future , revealing a gap between Forecast Available Capacity ,
FAC, and Forecast Required capacity , FRe. Additional capacity can be
ordered now to be completed by the required time, thus eliminating
the gap. This process will roll indefinitely into the future and all should be
well.
The physical realities with projects taking either 9 or 12 years to build,
requires greater complexity . Figure 11.4(b) shows a policy of concentrating
on meeting demand. Forecasts are made for 9 and 12 years ahead. The 9-
year gap is filled first by choosing 9-year projects of the right size and
additional 12-year projects are also ordered to try to fill any remaining gap
at 12 years.
On the other hand, 12-year projects are likely to offer lower operating
costs, and Fig. 11.4(c) shows a cost-orientated policy of selecting 12-year
projects as far as possible and, as it were, filling in any 9-year corners with
9-year projects. Both of these policies depend on projects, or project com-
binations, of the right capacity being available , and, as we have pointed out ,
the projects vary considerably in their potential generating capacity.
The influence diagram

In Fig. 11.5 the system contains only negative feedback, the two loops which
are apparently positive arising only as the drains from delays and being
dominated by the true negative loops.
Figure 11.5 is a very aggregated view, and Fig. 11.6 indicates some of the
complexity in the model, using the Cuyo basin for illustration, though the
same ideas are used for the other basins.
The first idea is that the capacities of all 26 projects are stored in the
Catalogue Table of Hydraulic Project Capacities. The second is that some
of these projects, numbers 10 to 15, say, are projects on the Cuyo basin. A
second table contains the preferred order in which these projects are to be
taken: perhaps number 12 is first, 14 is second, and so on. This order could
be varied to test, within the constraints of hydraulic engineering, alternative
plans for Cuyo development.
Every time a project is selected from the Cuyo basin, the Cuyo Project
Counter is increased by 1, but that only records the number of projects
so far chosen, not which projects they were. The rest of the diagram
shows a set of counters which record that Project N is moving from
one stage to another. At the correct point, the actual capacity of Project N
is picked up from the table and eventually added to the hydroelectric
capacity.
Forecast
Required
Capacity
FRC ' Gop to be IIlied
~ ~ ~ by new projects
~
k ~ Forecast
! -----
- -1- - - Available
I Capacity.
FAC. allowing
I for projects
I already in the
I pipeline
I
'-- . . l - _.... TIme
Now Now+CDEL
A The IdealisedCase
RevisedFAC FRC
ClJlVe wtth - _ ,
9-year projec11 , eman.......
12'Y"';;'"
.. J'
added ' A
gap
Strategy :F111 the 9 year .~ ...9;';;; I
gap and make up any ~
u
.. -l." gap _ , - FACwtth
12-year gap wtfh a u '*=--=-~-=-=-=_:-:_:::-:-=-.- -- - I- - I ~~~~reody
12-year project
If one Is available I I
I I
I I
'-- l.-.._---J... lime
Now NOw+9 Now+ 12
B The Demand Orientated Straegy
Strategy : First fiRthe

12-year Gop and make
up any remainder by a
9-yea r project If one
Is available
C The Cost Orientated Sfraegy

Fig. 11.4 Capacit y expansion policies.
These ideas are expre ssed in equations of the following form.

First , the Cuyo Indicator Start Rate is generated from CYSTANDP, the
next Cuyo project to be chosen, number 10, for instance, and CUYOPSW,
a 0/1 switch to start a Cuyo Project:
Long Construction
Deray ( 12 years)
Long Term Medium Construct ion

Capadty Delay (9 years)
+i'0~~~g
I
I
I
f
J
, +
, +
I
I
\ 1:/..ordertngRate~ Capacity
.
Medium Term
+
j
Hydro Capacity
Completto-i Rate
I I Hydro_ +
\ I Capacity Actual
\ 1 In Pipeline Hydro
\ I I Capacity
\ I I I
\ \ I I
Project Ordering I I
" Swnch \ I
I \ + \ ... Forecast Hydro " I
I \ + Capacity over· +
I \ Planning,Hortzon
I \ I
\ ' /
\\ ', ........ Future Gap -
/"
\ In Electrtcity . - -
\ ProJect Supply Required
~ +~.... Hydro + Ac tua l
PoRcy - - - - - C"8':~ity+- - - - - Eleetrtcity
Planning Demand
Hortzon
Fig. n.5 Basic feedback loop structure controlling capacit y growth.
r cyisr.kl=cuyopsw.k*(cystandp.k/dt)
Project number 10 is now stored for 1 DT:
1 cycds.k=cycds.j +dt*(cyisr.jk-cyerase1.jk)
r cyerase1.kl=cycds.k/dt
which is just long enough for the capacity of Project 10 to be selected from
the project capacity catalogue, TPROCAP:
a cycsds.k=tabsq(tprocap,cycds.k,0,26,1)
TABSQ is a special form of table function which gives the effect of a square
table (see the COSMIC User Manual for more details). The 26 in TABSQ
is the total number of projects in Argentina.
If this approach of using a control which lasts for only 1 DT is not used,
the same project will be injected into the system over and over again.
Developing A rgentina's electricity capacity 319
SWItch
~---C uyo
system
L
Cuyo
Project
Counter
.>:
+ Cuyo standing
Tobie wIItI
Curo proi«;t
1+
Cuyo projec t +
+
Erasing
~~ Ind ic ator -------01 Code 1
ora.r sta rt ra te (CYERASE1)
~1~n ~l
Delay !+ +
Cuyo projec t
Cuyocode
project design ca pacity starting
to constructlon design stage
tran sition rate
L
Cuyo capacity
starting design
stag e rate
Cuyo project capacity

/ ending design stage
c:.~.
hydraulic
ptO/«:t
1 +
Cuyo capacity
ending design
+
1 Cuyo capacity
In design stage
copcx:IIIw stage rate
Cuyo Cuyo c a pacity

Construcflon In construcflan
Delay pipeline
I I
+
Cuyo c apacity + Hydraulic capacity
c onstrucflon - - - - - - - " Installed
completion rate
Fig. 11.6 Influence diagram of Cuyo capacity expansion.
The full detail of the problem is shown in Fig. 11.7.
The model
POWER1.COS. runs a reference scenar io in which tables are used to
define the proportion of electricity capacity from thermal and hydroelectr ic
~ :,!,
~~ A
"- -:---~~
=~~~ :;~~= :;f~~':
Activatlng the sWitch 0 1
the c k»esf combinahon ----to-
""OOP
---- -r Establilhing the ~
"-r-
bt~ the mnmum
1
Estot::*sNng the mrinum
.\
~\
dltfefence aftet einW'latTlg dIf'I~ art_ ermnarng df'fet8OCe otte, eln'WlOttng
lI~~4 ~~1J~I- ,.I J~':t jT~T

C u yo NeuQuen Plato Patagonia
lOOO= c~Ll- .Ll.l, Ll.;..
~. i~~c~~m ~.-r\~ . Cuyo Neuq~ Plato PotoguOlO
p IQ IQ Ptoto Cuyoond Neuquen Cuyo copcx:lIy capacity c apacity capacity
~o str:::bv S1)ooncftlgby Neuquen Sloncf.by StorwH~' wponalOn elPOrnion eapanSlOfl 8llpon$lOll
PtotectI
Con-o'
Project
C0ft"05
Project
con-o~4long ' ..m _....
-,"""
PrOfed's
COft"O3
Pto;ect
COft"O2
Protect
COft"O I
TTTT
( l-gaP)
- j
~ ear_o'
~~
~~~~~ftr::~....... OIff 1 .. _ OI ~~~~t::n~~~1OfIS •• --=

ttloMtobovemediUm 'etmOOPL OIff 2 ....· . _ _ _ _ _ _ wittlmedlumtermgap
CNff3-----
Fig.11.7 Overall structure of the hydroelectric planning problem.

capacity. Prices are in 1978 US$ with a policy of maintaining constant

real electricity prices. Note the extensive financial sector. Make some sen-
sitivity tests. Add print statements and carefully check that the model works.
Figure 11.8 shows that the complicated equations for hydroelectric capac-
ity addition produce the required discrete additions of capacity. The hydro
sector produces growth in line with the intended plan. The trick of adding
the capacities from the four hydroelectric regions gives a convenient display
and the large steps in capacity are some of the very large projects from the
Plata basin coming on line.
POWER2.COS, tests policy options with various scenarios of demand
growth and oil price . With moderate growth in oil prices and low growth in
electricity demand capacity, Fig. 11.9A, is added to provide at least the
planned margin of safety, and return on investment is, if small, as least
always positive. Depreciation cash flow and after tax profits always exceed
the rate of repayment of long-term debt and the ratio of debt to assets is
kept under control. In short, Argentina's national electricity company is
solvent and can meet the demands imposed on it, under these particular
conditions.
The lack of robustness in the policy of attempting to maintain constant
real electricity prices and expand to meet demand forecasts in Fig. 11.9B,
produces a financial catastrophe late in the 1990s. Other cases set up as
reruns in POWER2.COS produce much the same result. In short, the
official policies are not going to work, mainly because they were never
designed from a dynamic point of view.
As with the New Zealand case, there is much that the student can do with
this model. One policy option is suggested in Fig. 11.10, namely one of
gearing the real price of electricity to the financial needs of the electricity
company. Optimization will analyse this system in depth and design rich
policy options for it.
Other options for improvement lie in the relationship between the elec-
tricity supply and wider national economic growth".
Summary
This model and the underlying problem have been described in somewhat
less detail than the New Zealand case, mainly to stimulate you into investi-
gating the model for yourself.
Technically, it is very complicated and shows how it is possible to use the
concepts embedded in a sophisticated system dynamics software package to
go beyond simple continuous models.
12 System dynamics has been applied to the modelling of the growth of developing nat ions.
Saaed (1994) is an excellent example. For an earlier study, see Forr ester (1973).
'NormoI'
ElM:IrIcIty
1'rIc.
Debt
Wrltlen-down
Volueof
.i-> RO\ftO+
Assets
~
\-: 8ectrclty
~~j
~-.
1.Rate
~~~- Balance
Fig. 11.10 Endogenous pricing policies.
COMPARISON OF THE NEW ZEALAND AND

ARGENTINE CASES
These two models are very different and the reader should make a compari-
son between them. For instance, the New Zealand model is only about half
the size of its Argentinian cousin. Its equations are simple, whereas the
power problem is about the most complicated the author has ever worked
with.
. Is the Argentinian model 'better' because its techniques are more
elaborate? Is the New Zealand model 'better' because it is easier to under-
stand? It could be that the New Zealand model is 'worse ' because it
oversimplifies reality too much, or perhaps the Argentinian model is
'worse' because it has used overcomplicated techniques just to look
impressive.
Such comparisons are virtually meaningless. The only way of really judg-
An air traffic control system 323
ing a model is relative to its objectives" . Review what was said about the two
models and then try to evaluate them.
QUALITATIVE SYSTEM DYNAMICS
We now turn to some instances of the use of influence diagrams as models

in their own right. The purpose is both to show the scope of qualitative
analysis in and to provide opportunities for the student to think about
problems and discuss them, perhaps in class or a syndicate. For that reason,
some of the cases will have very limited explanation and you will have to
work things out for yourself, seeking feedback loops and thinking about
what they imply for the system's behaviour. The student may find some
interesting opportunities for quantitative modelling in these studies, mak-
ing assumptions about numerical parameters.
The case studies cover air traffic control, drug trafficking, the mainte-
nance of business effectiveness during a reorganization and the criminal
justice system in Britain.
As we work through these rather disparate problems, we shall not always
adhere to the strict formalities described in Chapter 2. Influence diagrams
are models, and all models are tools for thinking with. The emphasis is on
the thinking, and the tool should be adapted to promote thought, not
followed so strictly as to inhibit insight.
In studying these applications it is important to bear in mind that none of
them required more than a few days of effort.
AN AIR TRAFFIC CONTROL SYSTEM
This problem (Fig. 11.11) is, at first sight, one of physical flow. The essence
is that the more aircraft there are in the zone relative to the ability of the
control system to handle them, the longer it will take for aircraft to get
through the zone (we have all been in an aircraft which is 'stacked'). In
practice, airlines negotiate with each other for landing slots, thus hoping to
ensure that aircraft can enter the zone when they can be accommodated.
The weather and problems at other airports ensure that the plans do not
always work out.
There are two pressure points in the system. Buying more ATC comput-
ers, radars and people will increase Work Capacity, but there may be no
point in doing so unless the Air Traffic Configuration is well designed. The
+/- sign on that link is to suggest that a given change to the configuration
may make matters worse or better, depending on what it is. Obviously, no
ATC planner aims to make matters worse by rearranging flight routes but
IJ These sections are derived from two Ph.D. theses, but describe only selected parts of those
studies . One should not attempt to compare the two dissertations, merely these two models.
------ .....
//
_.luOle ,/
S;:::::~~ ,/
-.;. /
• I
: I
~ I
Aircraft Requiring
Entry to ATC Zone
.-
Fig. 11.11 Overall influence diagram for an air traffic control zone .
Air Traffic Control
ConnQuration
(Pattern01Routes and
Spacing Rules)
I
I
I ,.....----- .........
I ~ <,
t +/- ,/ <,
Available /
<,
Loop B -,
~lace /
Loop A \
+ y II \ +
Aircraft Requiring I- - Rate of Entry ~ Aircraft In +- Rate of Leaving ~ _ Tr~nsit
Entry to ATC Zone
to ATC Zone ATC Zone ATCZone Tlr
+ t \ ...
:I \ Loop C /
I I \
I , I /
\ +..... '" /
\ Work Load for - - - - - .... /
\ Air Traffic Control /
/
\ I /
\ I ,.- /
, I ./
./
" - t .- .-
.... SpareATC
Work Capacity ------
+A
I
I
I
Worlc capacity
In Air TraffIc
Control Zone
Fig. 11.12 Loops for an air traffic control zone.

dynamic systems can sometimes behave in counter-intuitive ways, so it is

unlikely that all changes will be for the better, and careful design may be
called for.
In Fig. 11.12, Loop A should regulate admission to what the system
can handle, sometimes achieving its effect by delaying the departure
of aircraft from, say, Paris so that the London ATC Zone should have
capacity at the expected arrival time. Loop B, being positive, ensures that
problems are likely to get worse once they arise, though Loop C, acting
as a drain from the system, ensures that the problems will eventually
dissipate.
Figures 11.11 and 11.12 are somewhat in the spirit of the DMC problem,
in that the emphasis is on the physics of the system. Air Traffic Control is,
however, but part of a system which also involves the airlines, the passen-
gers, the airports and the aircraft manufacturers.
Some of this complexity is shown in Fig. 11.13,which is probably at about
Level 2 or Level 3 in the cone. Numerous delays are indicated, of very
different magnitudes. Since the problem is now essentially one of behav-
iour, no physical flows are shown , though they would be present in a more
detailed diagram.
Ideally, the mismatch between supply of and demand for ATC services
in the upper left-hand corner will trigger investment in ATC systems to
increase the ability to handle flights and redesign of the configuration to
get the most out of the available space, though the investment may be
constrained by the vast cost of ATC systems. ATC charges may have to
rise to recover the investment, the countervailing pressure being that
charges can be spread over more flights. It also triggers efforts by the
airlines and the ATC managers to rearrange flights to what the system can
handle.
The economics of passenger and airline supply and demand are indicated
in the lower left-hand corner, as are the effects of traffic growth and the
changing size of aircraft.
Several pressure points are shown . One , overflights, is largely a matter of
geography; the others are under the control of the ATC authorities, the
airports, and the aircraft builders. Which is which, what do they mean and
why do some have +/- effects?
THE DRUGS 'PROBLEM'
The use of drugs has been studied by several system dynamics workers.
Homer (1993) developed a quantitative model of the dynamics of drug
usage but to do so requires assumptions about the effects of, say, the supply
of drugs on the rate at which people start to use them, Homer making
diligent efforts to fit his model to available data.
Figure 11.14 attempts to capture the initial and tentative thoughts on the
/ _- - --00 - - - - _ - +"1~6~~ - - - - - - -
" .... .... IqriSiTO§l
+. 1 " " ,,/
Pressureto 0.
Reduce or » <
,, I + .. '
SrooothDemond
I \ \ // " ."
- - - - - -{};E - - - - - - - - - Investmenlln
-,
I ' I +" -- I I AIC ....
....
I ' I Pe<eelv~SuppIy .c -,
I \ I 01 ATC5efVices ,v,
\ \ I I A+/_ + I
,,
+ \I _ "I I if \
Chotac:tetIItII _ _ . . £""- _ _ ./ I Ale
\\
A_
, __ - - _ _ - I WOf1tCcpodIy \
I
\ "".... I " I / I
\ "" I I I I
\ +/_ Demand,or ATC~_ I I I
\ ~C 5efVices , I /
\ II + I ,," I
I
\\ // _ _ ------ - - - - __ ' I // " I
\ I/""'" / ....... . . . . + t +/- ..... ""'- +/
'\, I ,/ ..... Ret e o f _ ..... _ AICChor ges ...
' . . Flights " " OtIetfI/gtft / - - Handling Fllghts- + -'" .......
- -c,
+ / " SCheduled ,
" ,/
I ' ....... - - - - - - - - - -
-- -+/_
........ PrIcing
-, PolIcy
/ " Av_eA..,_ .... ""-AIrer+~Requlrhg <,
I . JIm -,
lD I I \ " .-- G round HondI~ -,
\8 I \ / <,
,,
I \ /
\ , I \
' , -IJ I \ I _ Po soenge< - \
....... Inte nded + \ I /' Sottsfactlon" - ._-- -- \I
<, \\:lIumeo'... -9-_ ' I ~p .... <, ,,"
AHra1llc A - .... ' I " / I ,
/ " , , + 1+ / , / \ I
/ + ," • ~ + \ I
/ - - Passenger \ I
,. / ..... -....,NlJllbers - - - - - - - - - ... Dek:Jysto \ /
II ' I
Growth
ITr~1 /
~ / I
+~
~- ' I
-, + /'
I
_ _ _ _ _ _ _ __ .. Achievob6e ,/ It' E2 ,
\ \ .................... + AirUnEf /
/ ./ Prices ~- I \ . . . . - - - - - - - - "'c osts
/ I I \
I \ I ' ""
Number 01 \ I ' ....- - "
Airlines C ompe" ng \ I HandIIiIf1 -- "
+'\ \ I - -- "
'.... _ .......
" + +I
.. Air1tne" --
- __ 8 c --PrOttlObItty - --
- ------- --- ----
Fig. 11.13 The context of air traffic cont rol.

+
___ - - - - - - - - - - -W>O~ ce ENOl" __
--- <,
--------~--- - -----------
/, ,/'" - - - - - - - - - - ......-;.'.:::::: _ __ '\. _ _ _ _ 1'oIiee
,/ ..... _ ,. _..... \ ..... _ EtIec".,.,.." • _ _
// +,.....,/ ... - > - ....... ....... \ +// +

/ MOrket.----------,.....::::_ <, -, +~p
/ / Soturatoon _ / / - - - _ _ -, <, lasseS
,/ / -- -, S.. .tilyol
I / / ...... ::. ......~-----' , s.nt.ne••
I / + I + • ............ --..... "- + /
,
J I t PrOduct Range ...... <, ..... ..... " - ~----
, Dtug and Make!.ng '\ -, _ :::: 'Drug
I: "" " - - PrIC~ + • Enort \ ",:-y / - - Tratltoc.~efS+
l
I , / " " I / -, +
\ / \ , I / -,
\ I / \ , I / '\ ~+
\ \ / _ - - - \ - - - - - - 1__ I / " Returns
\ +~ t + _----- -, "1 ---_,_ I ' . _..... ,
\ Drug -related" _ , , ' I v- -; \ Recnsts ..... ~
'v Cnrne bv users + "I J I - _ ...... +" \ - - - - Eft"".
and TrOtflc kers '....... I I I .......... \ I
I " , I II ' ,\ I
I 1- I -, \ I /"
, \ .......... .:.
+'"++ -, , \ 1 /
, Non-users ..- t ~ransihon '\ I I - otnc kers
l ..... - " Rate • users -e Drug-related Derating
Cos ts
\ // \ -,t , \ _ + I'\ \ + deaths
+ 1-
\ , \ ' ~ +
\ I \ I' 1 \ /1
, / ' I ' / 1 I
\ ' / ' I - - - - - _/ I I
\ I )\ " , Eflecf,._d I I I
\ I '\ - - Anll-drup - I
\ \ E~g~~~~ -, . . . . ActionJl 1me I
I
, \ Actions -, , 1
', \ t ' , I I
,, \- / + . . . _- I ,
I ." I
< , - +. per~,::gr~OI /'./ -,- ---
5evooly of /
Problem /
/
t + /
\ /
/
\ ./
, ./
'\
<,
-_/
Fig.11.14 A tentative model of the drugs 'problem'.
The drugs 'problem' 329
problem. The essence of building this diagram was the entity/state/transi-
tion method once one recognizes that there are three entities (at least) : the
ordinary population who are , or are not , drug users, and who can be viewed
as moving between the two categories, the population of drug traffickers
and, of course, the physical flows of drug s.
There are many areas in the model where questions are posed rather than
answered. For instance, Homer argues that the price of drugs is determined
by the traffickers ' operating costs, rather than by more obvious forces of
supply and demand. Perhaps excess supplies of drug s (the Drug Stocks in
the diagram) stimulate traffickers to extend their product range . The influ-
ence of seizures of drugs on attempted inflow is shown by a + /- sign to
suggest that up to a certain point , seizures increase attempted inflow. Pre-
sumably , the intention of the enforcement agencies is to seize so much that
the sign turns negative and the trade becomes so unattractive that it dimin-
ishes. Perceptions by the non-using population that there is a problem may
be a driving force in increased efforts against the trade.
An ideal way to make progress on a problem as complex and uncertain as
this is by the study period method, involving discussions between interested
parties and examined in the next chapter.
It is often useful to simplify a model to its essence by moving up the cone
of diagrams . Figure 11.15 is a Levell interpretation. It suggests that the
perceived severity of the problem triggers efforts to address it while, at the
same time, other forces increase the drugs business. Two sorts of govern-
ment action are shown: those to ameliorate the problem by social and
educational means and those to destroy the value of the trade by preventive
.
Perceived +
---- ------:~~~ _ "
/
/
/ - ""
'" "
" '"
",/
I \\
I
I " I
I \
\ Drug -rela ted
I I \ / Deaths
+1 I I
Awareness Non.Jsers ~Transitlon
Rate ~ I __
Users -
ot Pro b le m
II _+t+ --- ,
, -,
11
II I
I I
I ,
\ \ Government I \ '\ Government
1 \ Social Policy I I \ + Preventive
' " + Instrumenfs and
\ ..... _ .. scoie ono /
"," \
\
'Va lue ' ot _
Business .. - - - - - P~~I:~d
and Punitive
I Effectiveness
\ I Effect iveness ot
\ o~~~es \ Trafficker / Resources
-, , Product Range '" /... Dep ioyed
- - and Mar1<eling .- - - -
4,"
~
"\ Strategies + I +
'
"" '""
\
' 'Government Strategic __ - - '"
<. '--~~s----- / /
/ /
-- - ---
Fig.11.15 The drugs 'problem' at level 1 in the cone of diagrams.
/' -
/
/
Negative I
Non-users +- _Tril~on _ _. . .
-/ +
/
./
--
Government 5fratfIQic __ -
Bala1JCfl artd _ _ _ - - - -
Public Attitudes
--- ---
Fig.11.16 Feedback loops at level 1 in the drugs 'problem'.
and punitive methods. The government must strike a balance between the
resources allocated to the two approaches.
The effects of these mechanisms are highlighted in Fig. 11.16, which
shows two of the loops. There are more and, armed with an understanding
of them , the reader may be able to make more sense of Fig. 11.14 . Certainly,
an attempt to improve Fig. 11.14 (which is not the same as making it more
detailed) would be a worthwhile project.
BUSINESS EFFECTIVENESS - THE CASE OF

UNIVERSAL INDUSTRIES14
Universal Industries has expanded considerably over the last 10 years and
its staff are scattered over offices in the south-east of England. This causes
much inconvenience and makes it difficult to reorganize the activities and to
take advantage of modern technology such as the 'paperless' office. Ac-
cordingly, Universal intends to move all its staff to a single large office in
Shiretown in the English midlands.
The move will take a year to complete and there will be a need for
staff to commute from London to Shiretown and vice versa for meetings.
This effect will be largest when the transition is about half complete. While
the move is going on it will be difficult even to know where people
14 The case studies in this section and the next were provided by Jonathan M. Coyle. The
author has somewhat simplified them and adapted them to the style of the book.
A study of police force measures of performance 331
are (internal telephone directories are usually out of date before they are
printed). As the move progresses, more time will be lost in travel from
Shiretown to Universal's customers and suppliers who, by and large , are in
London.
While individuals are moving home and office, they are effectively un-
available for work. To that effect must be added the fact that there are a
number of people who do not want to move at all. They may be expected to
increase wastage from the London staff and to add to the time lost due to
illness. Naturally, wastage must be replaced, which will add to time lost in
recruitment and training.
In Fig. 11.17, the Work Profile represents a base load with fluctuations.
The Actual Work Output is based on the number of effective staff and a
normal working week, though the introduction of the 'paperless office' at
the new site will have the effect of increasing the work output.
A significant factor is the nature and extent of the departmental reorgani-
zation. This is shown with a ? sign (another example of being flexible with
the conventions of influence diagrams) to show that the effect on the ability
to locate people mayor may not be beneficial, and will depend on the
nature and timing of the reorganization.
The model is based on a very simple idea of it. 'Business effectiveness'
can be a very vague concept, in this case it has been represented as the
fraction of the work output which can be achieved against that which is
required. To that, of course, could be added the financial costs and benefits
of the move. The object is to come up with a Movement Plan to minimize
penalties and maximize paybacks.
A STUDY OF POLICE FORCE MEASURES OF PERFORMANCE
This study is an example of how the qualitative strengths of system dynam-

ics can be applied to address a problem in which the boundaries and issues
are far from clear. The diagrams presented here were intended to be purely
illustrative and no claims can be made about their comprehensiveness. The
diagrams do not represent United Kingdom government policy or opinion
in any way.
The study was conducted in two iterative stages:
1. A short period of initial fact finding followed by the creation of a draft
influence diagram.
2. Presentation of that diagram to the client followed by revision of it.
This study also shows the flexibility of influence diagrams. In the drugs
problem, the major diagram was presented first, leading to two higher level
portrayals of the major elements of the problem. In this case, the approach
is to build the diagram up in stages , so that each can be understood before
the next is encountered. Dotted lines are what was on the preceding
--- ------- - - - --- - -- - - - - -
/' --- - - - - - -~~~~~~~~-~~,~~-
<,
/
,/
- .r
---- + / ---....
<, "
--
/ -7~----- -- -- - ~oo~e ..... , , <,
<,
-- / <, -,
/ /' - / Movement <,
~ <, \
/ /' Plan -, \
-,
\
r /
/- "' - -, \ \ \
/I Pe ople + + De~rture - Pe opl e at \ \ \
/ I ' setne d In 'at .. se:~~ •.., .. DMOVf Rate -- Old Offices +" \
/ I New Office ~ I fra cti o n. o r \ \
\ / Sta ff In
I
/ I
I
\
-, - op e MO\llng
J + / ; ,.
New O ffic es
I I
,\
/ se ttllng ln / /' / \ \
I \ -, I
--' "
\ \ I
1 \ , ' /' / _ /' /1 \ \ I
<, ---/ »> /
\ '..... »>: _-- /' I \ \ I
II \ ..... + . + _ _ _ _ +/- /' I I I I
....... - - .. Tota l EHec tive..- - - - - - H L t /
I \ Normal Sta ff - - - - - - - C~~~u~~g ... I I I I
I ' Av_rog _ I~ _ - - BelWeen Offices I
-, WorIClng" I \' /' -' " + +/_ / I I
\ '- , W._k " " + ++ \" /' I ime to
I / ' TLocate /
ok' I I
\ <, ..... ....... Effective Nom inal \ ' /' I ...../..... Pe.Ple I I
\ - - - Working + Work Outp ut , < Commuting..... /
+~
.. Time per - - - - ~ (hours/week) \ / , nm_ _ ..... \ ? I I
\ /+We ek \ 1'- <, <, + »>: / \ / I
\ // \ r >; - - - - . Hour slost in _- /
Internal + / <; _ _ . ' II
\ \ / ' ,- _ _ - - Co mmunications + /'
\ +" ., - - -' , ok'
I
Extra Trove'
\ Act ua l - - - -.... + Hours l ost Time 10 I
- Wol1< .... - - - - - - .... ln O the r .. - - - - - Cus/ome"
\ _ -", Ou tput ..:::. _
~ _ ./ Tra ve l + ond$up~~n
I
-, ....... ftom New Olfice /
........ + Tim e l ost _ - ---- \ \ --- -.::: ---- ---- /
- -. t~~~~~nt - \
/
-, -.. -.... Hours lost /
t o Illness ...-_ Peop le MOVing /'
- W~~~~~'t /'
,-, +
Mo"e
I Wo rk Profile r-----_-:. Oesired Wo rk

O utput - --_
Fig.11.17 The 'business effectiveness' influence diagram.

AinIaI
~
--
~~[
_.~~ Time
~~
PropClllton of "'-. ...
~--------+!--- c';:'~~~tlan
C_R~\.. ~p=~
~ \ +
\ arne
":~~..[ .
. / v~'=-
++t~
~~~~~~
Fig.11.18 Overview of police performance (crime occurrence and prevention).
diagram and solid lines show what has been added. Factors which might
appear to be parameters in one diagram will be variables in the next. The
true parameters will be shown in bold italic type in the last diagram, Fig.
11.21.
Figure 11.18 represents the first element of the diagram.
Volume of Crime represents a function of the number of active criminals
in the population, the number of crimes each criminal could commit if
there were no police, and the time between actual crimes. It attempts
to capture the idea that, since a car thief's 'productivity', in crimes per
month, may well be higher than someone attempting to commit a complex
fraud. However, the Perceived Volume of Crime is a function of the
Proportion of Crimes Reported as well as of the actual value. This means
that the complex fraud described above might be seen by the public to be
more significant or worrying than car crime if a higher proportion of fraud
cases than car crimes were to be reported. In the late 1980s this was indeed
the case, though, in the early 1990s, car crime has higher visibility than
fraud.
Perceived Volume of Crime causes an increase in demand for resources
to be allocated to Police Visibility (police officers on foot patrol, as opposed
to less visible car patrols) and also decreases Public Reassurance. In this
way, increases in the Proportion of Crimes Reported have two effects:
• They cause Perceived Volume of Crime to equate more closely to the

actual value and so allow the resources allocated to, say, crime preven-
tion, to match the real requirement.
• They increase public concern about crime, leading to a greater demand
for resources to be allocated to Police Visibility, which might not be the
most beneficial use of police resources.
Figure 11.19 expands the previous diagram to cover crime detection and the
Crown Prosecution Service (CPS). It suggests that increases in Perceived
Volume of Crime and decreases in Public Reassurance will also cause more
resources to be allocated to Criminal Identification. After a delay, the Time
to Prosecute, which is the time taken to prepare a case, and is constrained
by the Capacity of the CPS, this will increase Successful Prosecutions which
increase Public Reassurance about crime, thus forming a large and signifi-
cant negative feedback loop, running from Police Visibility via Response
Time to the volume of crime variables. The Proportion of Crimes Reported,
as well as affecting Perceived Volume of Crime , may also cause a decrease
in the Time to Identify Criminals and so increase the rate at which they are
identified. The Time to Identify Criminals is also a function of the Probabil-
ity of Arrest, which may increase as the Number of Crimes Reported rises.
Decreasing Police Response Time should increase the Probability of Arrest
and so enhance Public Reassurance. Unfortunately, Police Response Time
,_..... ~r::~r~~
of hIonn Procu. + ...O lrrl ll')Q!:> = -- - - - - - - - - - - P,OlxJb "11' _ <, <,
".~~t ct A"."
-, /
C flfl W\OIS + I
/
/ - - ' ......
-, ~ .+
I
I ~_ I ~pon,~
'"~
P.:.ob<:e
----- " '\ +
I?elu rnwgto --.... Nurnl::>efof OfCn
Acito' me'(
'''~"", / \ P,o ;<,mlty ot
I l ln~
I -'\ Between +
PC~Uflo,on/J
.~" e:i\m:c" I I -, ~ .-/ C """, ,", ,
( <,
/ I ". ~ ./ / +t <, ' ... .....
c..... -
~1lson ~ \ , . > I ( "
CC po<. li't' CoparcIPy \ '..... .:« J ... ...... .......
u,,,~a,:, / \ - - - - -- - -- I <, H flCM::tl,CY C:'
\, CUrT~ P!tJvel) l()n
C IIITlJr O IS
ffl:lolToOVl:!d fl o m
Popucmo ....
-I
h"" lo
PIO$tK:v1e
' ......\
\ -
......
t
+
<, ........
"~~~
OI C(~ed-
""" ......- --
I t t '\ , " ~
- - - - - - ---~ -- -
,D
II ('"
'\ \
T,n,el :)
bll corr *.l
used tc
Cllm",
.. ., / _- 9 - _ " "
C llrnr. a l +
:J \-.. . . ' , . ...... ...... .... / ./ ' '"
+ •_ _
t - ioe ntsc c tc n - - Pec otded <, , < , 1/ Publoc ;'-/ /
+ C IIIT\l:t --- ' Polic e l~eouulallCe ~
.... Vl!lbttr"r'
~ Sl-C(;e'Aru-
_ Prosecuncn s +
~ + +t "', _....... / // +~-
\L \ E ff~r::ZIOI I , ...... - "
Idenfll ic a hon 1
t I
I
I
I
Eftl(· le nc.yof
Pe sce l/ os.lblllty
~----, .- - --------
Fig.11.19 Overvi ew of pol ice performance (crime detect ion and the criminal justice system).
may be increased by diverting officers to Police Visibility. In this way, we

see that there may be a conflict between two of the control mechanisms in
the system (the police visibility loop and the response time loop) 15.
This is a good example of one of the major strengths of the influence
diagram approach. Police Visibility and Response Times are two variables
that the public often use to measure the performance of their local police.
However, the diagram shows that these measures conflict with each other.
This highlights a potential weakness in treating them as independent
measures.
Figure 11.19 also shows that an increase in Successful Prosecutions will
lead to an increase in Criminals Removed from the Population. However,
after serving their sentences, these criminals will return to the population,
unless the reform process is effective. The diagram shows that the criminal
identification process is a negative control mechanism , but it requires in-
vestment in prison resources to keep Prison Capacity Utilization to an
acceptable level in order to sustain the process. However, the link 'from
Successful Prosecutions to Public Reassurance may well be quite a strong
one.
Figure 11.19shows that increases in Actual Volume of Crime increase the
Proximity of Crime, the feeling that crime takes place 'close to home', thus
causing Public Reassurance to fall. This may have a stronger effect on
Public Reassurance than Perceived Volume of Crime. This last variable is
also affected by the rate crime is recorded. Two variables in this system,
Recorded Crime and Proportion of Crimes Reported, potentially distort
the Perceived Volume of Crime. Given that this last variable is the driver
for many of the negative loops in the system, it is clearly important that it is
as close as possible to the Actual Volume of Crime.
Figure 11.20 examines community relations, police morale , behaviour
and training . The parameters are now shown in bold italic type. Some of
them , such as Time to Become Used to Crime, are characteristics of human
behaviour. Others, such as Police Training and Crown Prosecution Service
Capacity, are pressure points which might improve the system's perfor-
mance . It ought to be possible to apply the techniques of constrained and
free optimization to deduce broad priorities for improvement strategies, as
was done in Chapter 9.
Improvements in Police Morale and Police Training will increase the
Efficiency of Criminal Identification, while Police Training will increase the
Efficiency of Criminal Identification, Police Visibility, and Crime Preven-
tion. These links are not negative : making morale worse or having less well
trained police officers is unlikely to improve their performance. Police
Morale is likely to be enhanced by Successful Prosecutions and reduced by
Complaints Against the Police and by those that are upheld . Complaints
15 You will be wasting your time in studying this diagram unless you are also finding and
assessing the effects of these loops.
Tlne fo - _
Ide n h!'t4 - -- - - - -
+ CllIT\II"ICIls I'oIwtIIaI -- -
oI:='~" ,'fI \ ~ <,
- ProlxJl.>l~"" - ,
1/ I \ IN' ~ + " r AUtnt ..-- . /
_. - -::;-:::::._ - - -,
I / _,- - -4 _ , , ... .... . . .
- I / I \ + .+ - - _ _ _ - - - - - - - - _ -: ~~!>d
,. / \ c fuo l l,lQ1urne _
Cllnl ~ts -+ / I \ I ofQmu / _ _ +Iimtl \
' Ift.tlutn ng 10 - - - _ _ .. N~er o r I , ~ .. + ,/ TITlt:'J ,-
/ Popu bh on Actve Crrnnals \ I t -" " "" Batween + ,/ \
\
+~ \ \ / \....... .... /' ..... ./ O lml t5 ....... ..... /
\
I
/ +. I
-..
/ <, - _L __ - ( I \. .... _ _ _ _ ..... ,/ +
t <, ....... ,/
\
- I I I I ....=.., \ \ \ " -- - - - .: " I / .... \
",:'0 lD I ....... \ \ \ ' .... » > I / \
5~~~~ I I I ~ ~~1On OI \ ' , - - - - I / ..... Ethc ....nc.:y o l \
, .. + I / I Cnm65 Report~ \ '~ + - - -- _ _ 1__ I CurTIeP ev en tlOn \
I \ I ,/ / - ,, \ -, -.. Pe rc OlYed _ - - - I - 1- _ + \
I \ I II - I .... ,. ~::: .... .... / I .... .... \
I ' - "~~";'';'''''I~Gm p,~o;;::,~. \ - + I t +\ ........ .... I I2l I I
k>pu1o,]11On \ \ 1\ \. . . . . ..... + ,/ I I
' 4+ I \. I -, "' Cmn6 / I
\ \ 1 \ ,, \ \. Preven tion I
I I \ -, +1 \ ' +t I I
I \ .... ~ljJ \ ' \ I I
\ I I \ I \ " \ I I
\ I I \ I \ <, , I ~ _ - 9 -_
" I / " cm,""':- , ', I 1/
\ \ / '\lII + _-- - ldentlflC01lOn - - _ ....
'r~~~ttd ' ., . . - ...... ....... ..... I I
\. \ ,/ ..f)- - - t+ + Pohce
\ - / - --
- - _ ";:VISibtllty
Succ"",.t .. - - - - I +
+
' ...._ ~'058C I.Jhons + \
t., _-,
1+ \. Elficl6ncy ol + 1 <, _ _ _ _ _ / /
J 1"oIIt»:~ -- CfNToIl'lOl '" I \
ftht*:~ Id en !lfl.::otc n I
II "'--..c. /+ I / \}
I 1
I I
I ~
I + Pc Ic e • Poec e
£~v~.ity
:I u~ ~~ , ~I~ - + +
\ + -------
\, .- - - - "'i + _ siOi; ~'n tOi'c~ • T __
~ Pu-posd 0l"l<1 VOI Je5
\ \. -, fhttPollco
"b"''''''' '_ - .-/
"
--- - - -- -- - -- - - - - - - -- --------- -- -
Fig.ll.20 Ov erview of police performance (community relations and police moral e, behaviour and training).
Against the Police will increase if there is an imbalance between the ethnic
or gender mix of the force and its local population. This variable will also
cause a decrease in the Proportion of Crimes Reported, which has already
been identified as a significant variable. Other factors which affect the
number of complaints against the police include Adherence by Police to
Statement of Common Purpose and Values (which presumably could be
increased by further police training) and the number of criminals identified
(thus the police could be victims of their own success in some circum-
stances). Finally, falls in Public Reassurance could lead to increases in the
number of complaints against the police.
The analysis of influence diagrams is sometimes helped by laying Venn
diagrams over them to emphasize sectors of a model and to identify sector
interactions. Figure 11.21 does this for some of the measures of effective-
ness proposed by the Audit Commission (1992). Each sector is identified by
a name in bold Roman type and an arrow to the boundary of the sector.
It can be seen from Fig. 11.21 that there is considerable overlap between
some of the measures of effectiveness. For example, those connected with
complaints against the police are completely overlapped by other measures.
This may be a desirable feature in that measures which overlap can provide
checks on, or alternative insights into, other measures.
The diagram addresses the issue of the quality of service provided by the
police in a number of indirect ways. However, the value of using complaints
against the police as a measure of quality is that the volume of complaints
could increase as the police become more 'successful '. For instance, an
increase in the number of criminals identified could in itself lead to a
corresponding increase in complaints.
The feedback loops that deal with crime prevention, criminal identifica-
tion and police visibility are all affected by how 'efficient ' the police are at
that particular task. In other words, the Time Between Crimes is affected
not only by the volume of resources allocated to crime prevention but also
by the Efficiency of Crime Prevention. This logic can of course be applied
in reverse; if it was possible to measure the Time Between Crimes (perhaps
by interviewing known criminals at the time of arrest or conviction) and the
level of resources allocated to crime prevention, the analyst would start to
gain insights into the efficiency of a particular police force in the role of
crime prevention.
In summary, the diagrams illustrate how the technique could be used to
structure debate about, and research into, this issue, perhaps through the
medium of the study periods discussed in Chapter 12.
THE COLLAPSE OF A COMPANY
In an earlier chapter we examined the case of the collapse of the Mayan

civilization. Similar dramas occur in business life and, to encourage the
reader to search the management literature we refer him to some seminal
PulH:~
E lh nk/ ~f' n lh" r Rc,p unw
Mil lim l:'
:
~
=:;:4"- - - - -- - - - - - -
.
---- ..........,
. ;.. ",
, - - _.- - - - - - -
"'...... .- , -t" :" "" " .....
\
"' - ; - - '. I \ ". ,
-=
0c';;:..
",
i ; . .......
,,, .
""
I _ - '"
--,, , -y---____ _.:. _.
" , ...._ 7' - - +\ - , '- »>:" _- _ - : ...._ •.
-" - - - ..:- : : " - »> . " _ ,
"""UN,,.
1 C.... . _
I" I . .- . ,
+
I
" '.....
c~:'"
"""" ""'"
.." '0- - - - - ~.:...;~:= I
:"" ,, ~' I
I
>AC~. . C~::.. _:..: . :._______ ~'''' _'
, I _'" " " ,
I I
,
'·; -- - ,," ~_. -. _. '. ':..
"" II I
~
_ • . ..... ...., ,,"
I, " C_
,"
I ' · i > I ' "" ,, . , .... , , I ", I I
I '- _ _ _ _ _ _ , ,( ....I" \
I I I , __ " I \ I "
" cr. - .... -~:\.' ,- - " " I ...." I I
I I I ' .... " I I I
-Po"" "" I I
I"
ID I ~
C_ " \ II
\ , - - - - - --_-- "
_- - ' 1
-- "- -----'----r- _.,, ' ,C~.,,<..,,u_~,..,..,..
1' .
: .~~ :I I
"" I
,, I
' I .
I ,.. . , .\ \ -, " + I
I ~"
.).. _ _
\ \ \ / =- :
..
II "
, c~·I I - "
" -.tp~;:;
c._:,:'_
-, r;J I I 0 _
_ •.~",,,,,, I \ "
O' mI ~ \- -
..._ • I \I.)hI\", 1
I"".. ,P~Ii~~', \
I ' \ '+',, ', ,,\ : I I
I
II \
\
~"" 1\ "
" 1\ " I I .' I
\\ I \\ \\ \ .I
' , . II I I
\ , , "
I I I \ 'y "
I , 1 ,, \ I
I I • I
I \ I I
I \ I
I" \ • I
"\I' '\
A ,, "
\ \ ...._ _ ..,.', ,.J1
I
\ \
•
.. ". I _:" ' .. . . I
' .,.t,
'-
\ -" I
SUCC. -....
1' .. -- - - -
__ B " ' . ,, -
_\. ".\ r \
P ft()MlC oJIoOJ' +
Il'le_ e Vc>' + "
I \
' o J
• /. , \ '
COInr l• inh _ " "
/ _ _.... L ....-" .... __
0 MIl .- /
.., _ _
.. _ _ _ _ _ L ,
_ _ _ _ \ \
' I" I , AK· itn' "'.,:; """ " , -, \ _,, _ .
Polk . I
I • ~~ ,
' , I ' , , I '>:
I' ' I ,
I " ' I
' I I \
1' I " -, ,I I I
I I " I
,
; ' ....
~I
" I
I \ \ f'Ck:_ _ I I
• - V T
+ / , I
'
""'~!~ "
I , + I
1 ,," I
, ,," ,, "
I _ " \ ,,"
I - - . _ _ __ ,," ",, "
I I ,, " . s. -: « " .{
:
»> I
\ \ .>": > :
\ \ __
\I \ , ..... """" ....... ~_.OO . I
- I 'if ~ .... Crin"O\
\ ,/
./
' -,
-- -- --- -- "::::'";;':~~ .. - _- -- _- - - -- " · 'Suh "..d '
./
\ ""
' ....
,, ,
-,. . ,..,.., - ---
"
-- --------------------- -- -
Fig.11.21 Ve nn dia gra m of sector
-------- ------ -- ------------ -
s in police measures of effectiveness.
work by Hall (1976) on the collapse of the magazine The Saturday Evening
Post. Apart from its business interest, this case shows the power of a small
model and the use of other management science techniques, in this case
regression analysis, in a system dynamics model.
SYSTEM DYNAMICS APPLIED TO DEFENCE ANALYSIS
Introduction
This section uses the open literature to describe the work which has been
done and to suggest reasons why system dynamics has been a useful tool for
defence analysis.
Coyle (1981b) shows many of the ideas which have emerged in this body
of work. It was suggested by General Sir John Hackett's hypothetical
scenario , The Third World War (Hackett et al., 1978). The model proved to
be surprisingly easy to develop and to 'fit' to Hackett's data. When the
initial conditions are changed, the model also fits a different scenario
(Bidwell et al., 1978).
The model has 300 variables and requires a few seconds to run on a PC,
whereas some combat simulations consist of 200000 lines of code, and
require major effort to run . This difference in size and tractability will be
explored later.
WW3.COS can be run to produce Hackett's scenario of war after a
period of warning and Bidwell's scenario of a surprise attack.
Qualitative analysis of influence diagrams has been very useful in the
defence arena. A paper written before the Falklands War analysed the
problems of sea control over a group of islands (Coyle, 1983).
Qualitative modelling was further applied (Coyle, 1984) to the relation-
ships between East and West in the context of the balance of interest
between Northern and Southern nations which had been suggested by the
report of the Brandt Commission, mainly with the intention of posing
research questions. Influence diagrams have also been used to model coun-
ter-insurgency warfare (Coyle, 1985). The main insight was that the govern-
ment forces should emphasize preventing insurgents from forming parties
large enough to attack sources of government power, combined with a
policy of controlling the availability of supplies to them.
Wolstenholme (1986) describes a quantitative model of a Red force
attacking a Blue defender; an extension of system dynamics modelling from
strategic problems to more detailed analyses of Blue's fire discipline and
Red 's attack tactics. He comments that:
In general the role of system dynamics [is to create] initial insights into
problems prior to detailed conventional modelling. This is .. . an im-
portant role.
System dynamics applied to defence analysis 341
System dynamics analyses the influence of information feedback pro-

cesses on the behaviour of systems. It was, therefore, natural to apply the
method to military command and control (C2) systems. Coyle (1987)
describes models which cover two aspects of the problem. The first is
an influence diagram of the 'theory' of command and control systems.
The second is a quantitative model of the ability of given C2 assets
to handle specified volumes of information. The output indicates how
quickly a C2 system can respond to sudden events, and, indeed, whether the
system will ever recover, how long it takes the system to cope with extra
demands, and how much query traffic would be generated by a sluggish
response.
Wolstenholme et al. (1992) describe the use of system dynamics for the
design of information systems in general, using two military examples. The
first shows how a proposed information system can be used to regulate the
movement of ammunition more effectively. The authors make the very
profound point that
the proposed information system must be evaluated in terms, not only
of its immediate impact , but also in terms of its potential to support
more efficient policies.
The qualitative sea control model was developed into quantitative form by
Coyle and Gardiner (1993), in a model of submarine patrols. The work
dealt, firstly, with the construction of a fleet of submarines and their recy-
cling throughout their lives through major overhauls, and, secondly, with
the ability of a fleet of vessels to maintain a given level of patrols. The
output from the first model indicates how frequently a fleet of a given size
would be able to put a certain number of boats on patrol over a life-cycle of
20 years. For example, with a fleet of four vessels, there will never be four
available, whereas with a much more costly fleet of 10, four will be available
about two thirds of the time. The second model represented submarines
being despatched on patrol but having to return if there is a mechanical
breakdown.
Another maritime model (Coyle, 1992a) deals with a fleet of aircraft
carriers attacking enemy bases while surviving against enemy attack. The
technical interest lay in treating carriers as integer objects during deploy-
ment and combat. These discrete variables must be mixed with continuous
processes of carrier sorties, damage infliction and self-repair, to some ex-
tent, by the carrier crews. A copy of the model appears on the disk as
CARRIER.COS16 .
16 This model has some apparent dimensional inconsistencies: for instance CACE is
the total number of aircraft available and is the sum of bombers and lighters. Dimensional
validity is a vital ingredient of a successful model, but it should be used flexibly where
exceptional circum stances arise. The author, who has a lot of experience in modelling, spent
many hours carefully studying the printed output to make sure the model was working
correctl y.
Unconstrained optimization methods were applied by Wolstenholme and

Al-Alusi (1987) to the Red attack versus Blue defence model described in
Wolstenholme (1986), a most interesting aspect being the 'design' of the
non-linear policy curves used by Red to control his attack. The optimum
curves turn out to be very different from the curves intuitively chosen in the
simulation of the same problem and between optimizations using different
objective functions.
Constrained optimization is a theme in an analysis of defence expendi-
ture (Coyle, 1992b) using an aggregated model of land, sea and air warfare
in a hypothetical country. Several different levels of defence budget are
optimized separately, and the results indicate the priorities for spending.
Clark (1986a) modelled the dynamic effects on finance of the expansion
of the US Navy during the 1980s. He draws on Cobb-Douglas functions to
estimate operating costs.
Clark (1986b) studied the effects on the stability of Central America of
US arms transfer policies. His model simulates the acquisition of military
capability and perceptions of threat in four nations.
Clark (1993) has studied the economic context of defence pro-
curement. He discusses a model of the effects of cut-backs in the
forces on the procurement of assets and, in turn , on procurement by
manufacturers.
System dynamics has been applied to a considerable number of real
problems by defence analysis civil servants and consultants in several coun-
tries. Topics have included combat modelling, manpower planning, equip-
ment design, equipment support management and many others .
Some of this work is in the process of appearing. For example, Coyle and
Richardson (1993) have used a system dynamics model to evaluate lessons
from the Gulf War.
System dynamics has enjoyed a considerable degree of application to
defence problems. The reasons are not difficult to see.
Warfare is fundamentally a dynamic process governed by the initial
forces and weapons (parameter values), attrition and reinforcement (physi-
cal flows), command and control (information feedback) and heavily influ-
enced by non-linearities. That is precisely the paradigm which system
dynamics addresses , so it is not surprising that the approach fits the problem
domain extremely well.
Further, system dynamics, by its insistence on capturing the principal
mechanisms in a system, is consistent with a top-down, strategic view of
problems, which accords well with the expectations and viewpoints of sen-
ior decision-makers. The clarity and simplicity of a well-constructed influ-
ence diagram appeals to their sense of logic and does not demand too much
of the time of very busy people, which engenders a feeling of control and
problem ownership .
A factor in the success of system dynamics has been the willingness of
system dynamics modellers to go outside their own tradition of rather
References 343
broad, continuous models and to tackle problems involving discrete events

and stochastic processes, while still aiming at the holistic paradigm of sys-
tem dynamics . The flexibility of the software has been a powerful tool in
achieving that aim.
Wolstenholme was quoted earlier as remarking that a system dynamics
model might be seen as a precursor to a detailed conventional model. While
that is certainly true , one feels that it is not the whole story . There is also
a role for system dynamics as a tool for guiding modelling strategy. If,
for example, a system dynamics model showed that the investment in a
new weapon system could not be beneficial for any reasonable range
of aggregate parameter values, there would be little point in any
further detailed conventional modelling. One would therefore revise
Wolstenholme's remark to read
the role of system dynamics is to create strategic insights into defence
problems as a stimulus to, and guide for, any further work which might,
or might not, be required.
That applies to every area to which system dynamics has been applied.
SUMMARY
This chapter has explored the application of system dynamics to real prob-
lems. The aim has been to stimulate the reader's insight into the capabilities
of the discipline and to show how problems of great complexity can dealt
with. The reader will get most value from the chapter by running the
models, studying their output, becoming familiar with how the y work, and
then experimenting with them . Optimization will deepen understanding
even more.
The examples , and the instance cited in the system dynamics bibliogra-
phy (see footnote 1 in this chapter) cover such an enormous range of topics
that it is hard to identify problems to which system dynamics cannot be
applied. That is not the same as saying that system dynamics should be
applied to those problems, and, in the final chapter we shall consider some
criteria for assessing the circumstances in which system dynamics might be
an appropriate methodology.
REFERENCES
Alfeld, L. (1975) Introduction to Urban Dynamics, Wright-Allen Press.

Audit Commission (1992) Consultation Paper: Citizen's Charter - Performanc e
Indicators, September.
Bidwell, S. et al. (1978) World War 3, Hamlyn, London.
Boxer, C.R (1965) The Dutch Seaborne Empire , Hutchinson, London.
Cavana, RY. and Coyle, R.G. (1984) A policy analysis model for New Zealand's
plantation forestry system, Ne w Zealand Journal of Forestry, 29(1), 24.50.
Clark, R (1986a) Operating and support costs of the US Navy: some analytic facts,
Budgeting for Sustainability, Military Applications Section, Operations Re-
search Society of America.
Clark, R.H. (1993) The dynamics of force reduction and reconstitution, Defence
Analysis, 9(1), 51-68.
Clark, T. (1986b) A management systems analysis of the arms transfer process,
Decision Sciences, 17(4), Fall.
Coyle, J.M. and Richardson, H.D. (1993) A recreation of the Gulf War using the
CARCO system dynamics model, presented at International Symposium
on Military Operational Analysis, Royal Military College of Science,
Shrivenham. (Copy available from Coyle at HVR Consulting Services,
Selborne House, Mill Lane, Alton , Hampshire GU34 2QJ, UK.)
Coyle, R.G. (1977) Management System Dynamics, John Wiley & Sons, Chichester.
Coyle, R.G. (1981a) Modelling the future of mining groups, Transactions of the
Institution of Mining and Metallurgy, 90, April , A81-A88 .
Coyle, RG. (1981b) The dynamics of the Third World War, Journal of the Opera-
tional Research Society, 32, 755-65 .
Coyle, R.G . (1983) Who rules the waves - a case study in system description ,
Journal of the Operational Research Society, 34(9), 885-98.
Coyle, RG. (1984) East and West and North and South: a preliminary model,
Futures, December, 594-609.
Coyle, RG. (1985) A system description of counter-insurgency warfare, Policy
Sciences, 18, 55-78.
Coyle, RG. (1987) A model for assessing the work processing capabilities of mili-
tary command and control systems, European Journal of Operational Research,
28,27-43.
Coyle, RG. (1992a) A system dynamics model of aircraft carrier survivability,
System Dynamics Review, 8(3), Fall, 193-212.
Coyle, R.G. (1992b) The optimization of defence expenditure, European Journal of
Operational Research, 56, 304-18.
Coyle, R.G. and Gardiner, P.A. (1991) A system dynamics model of submarine
operations and maintenance schedules, Journal of the Operational Research
Society, 42(6), 453-62.
Coyle, R.G. and Rego, J.e. (1982) Modelling the discrete ordering of hydro-electric
projects - the Argentinian case, Dynamica, 8(1), Summer.
Forrester, J.W. (1961) Industrial Dynamics, MIT Press.
Forrester, J.W. (1969) Urban Dynamics, MIT Press.
Forrester, J.W. (1971) World Dynamics, Wright-Allen Press.
Forrester, N. (1973) The Life Cycle of Economic Development, Wright-Allen Press.
Hackett, J. et al. (1978) The Third World War, Sidgwick & Jackson, London.
Hall, R.I. (1976) A system pathology of an organization: the rise and fall of the old
Saturday Evening Post, Administrative Science Quarterly, 21, June .
Homer, J. (1993) A system dynamics model of national cocaine prevalence ', System
Dynamics Review, 9(1), Winter, 49-78.
Macrae, H. (1994) The World in 2020, HarperColiins, London.
Meadows, D.H. et at. (1972) The Limits to Growth, Potomac Associates.
Meadows, D.L. et al. (1974) Dynamics of Growth in a Finite World , Wright-Allen
Press.
Naill, RF. (1977) Managing the Energy Transition, Ballinger Publishing Company.
Roberts, E.B. (ed.) (1978) Managerial Applications of System Dynamics, Wright-
Allen Press/MIT Press.
Saaed, K. (1994) Development Planning and Policy Design, Avebury Press.
Wolstenholme, E.F. (1986) Defence operational analysis using system dynamics,
European Journal of Operational Research, 34,10-18.
References 345
Wolstenholme, E.F. and Al-Alusi, A.-S. (1987) System dynamics and heuristic
optimization in defence analysis, System Dynamics Review, 3(2) , Summer.
Wolstenholme, E.F., Gavine, A., Watts , K. and Henderson, S. (1992) The design,
application and evaluation of a system dynamics based methodology for the
assessment of computerised information systems , European Journal of Infor-
mation Systems, 1(5) , 341-50 .
CHAPTER 12
Summary and conclusion
INTRODUCTION
We have followed a long and hard road to study the principles, techniques
and applications of system dynamics and have now reached the stage of
trying to draw the threads together. Accordingly, this chapter restates the
essence of system dynamics, and discusses problems of applying it in prac-
tice. That leads to some considerations of the assessment of the quality of
models. Approaches to the teaching of the subject are explored and some
further reading is suggested. Finally, the author indulges himself with some
personal comments.
THE ESSENCE OF SYSTEM DYNAMICS
Many aspects of system dynamics have been described in this book. The
techniques of influence diagrams and examples of their use were studied.
The principles of simulation were explained. Modelling techniques and the
formulation of models have been explored and the depths of optimization
were plumbed. In the preceding chapter we tried to bring all those topics
together in a survey of the applications of system dynamics, including some
rather detailed case study models. Those topics are all very well and the
reader will have expended much effort in mastering the material, but what
is the essence of it all?
To see the essence, recall that an aircraft flying from, say, London to
Hong Kong does so almost entirely under automatic control. During the
flight, the aircraft's weight changes considerably as fuel is used, it flies over
high mountains and is exposed to head and side winds which are largely
unpredictable. This standard of performance is achieved safely and eco-
nomically every day, and practically regardless of the weather, because the
designers of feedback control systems for large aircraft and many other
engineering systems know their business. They are helped considerably
by having access to a large body of accepted and tested theory of, say,
aerodynamics and some powerful mathematical methods of control system
design.
Socio-economic systems must also be, so to speak, piloted through time,
and we know from the practical experience of daily life that they are all too
The essence of system dynamics 347
frequently far from successful in that enterprise. The most dramatic exam-
ples are the decay of cities, the changes in social behaviour, the cycles of
economic expansion and recession and the rise and fall of nations. Less
dramatic cases are the growth and decline of business firms and the daily
efforts of managers to meet targets for production, to recruit staff or to
complete projects on time and to cost. Many other instances have been
mentioned in the book; even more will occur to the reader.
The essential aim of system dynamics is to achieve in socio-economic
systems the standards of controllability and dynamic behaviour which are
commonplace in engineering systems. The decades of work in system dy-
namics which were summarized all too briefly in Chapter 11 offer ample
evidence that these targets are achievable in a very wide range of socio-
economic systems. The paradigm, or fundamental point of view, is that
dynamic behaviour is produced by feedback mechanisms and that the qual-
ity of the design of the policies which regulate those mechanisms and the
structure of their interactions are the factors which produce good or bad
performance.
To be sure, the system dynamicist dealing with a managed system faces
more difficult problems than the control engineer. The first is often a lack of
any accepted theory of the problem in question. The saving grace is that all
socio-economic systems contain physical flows of something, and getting
the physics right is the key step to understanding a problem and getting a
good model. The injunction to think physics was emphasized when we
discussed influence diagrams , and the principle of physical flow is, perhaps,
the nearest thing system dynamics has to a general theory for structuring
and understanding problems.
Further difficulties arise because socio-economic systems usually contain
significant non-linearities, are often larger than engineering problems and
cannot be subjected to laboratory testing under controlled conditions. Even
more than that, socio-economic systems may well involve positive feedback
loops, or badly behaved negative loops, capable of driving the system far
outside the normal operating range within which an engineering system can
be made to remain .
Even further, socio-economic systems cannot always be represented as
objects which are separate from the environment that causes the shocks. A
Boeing 747 does not create the headwind with which its automatic pilot
must cope; a business firm, on the other hand, may very well create or
destroy the demand for its products to which it is trying to react. One
consequence is that system dynamics models of socio-economic systems
quite frequently produce counter-intuitive behaviour, in which the result of
a policy or structural change is different from, or even opposite to, what
logic might initially have predicted. Thinking more deeply, with the aid of
the insights from the feedback structure, should reveal the reason for the
counter-intuitive response, and that ought to enrich one's understanding of
the policy problem the model represents.
348 Summary and conclusion
To add to all that, socio-economic problems, because of their non-

linearities, positive feedback and sheer size are simply not amenable to the
rigorous mathematics of control theory. Indeed, one of Forrester's power-
ful insights was to see that simulation, supported by software with a suitable
syntax, was the only way in which socio-economic problems could even be
represented. The techniques of optimization of system dynamics models
take that idea a valuable step further.
If socio-economic systems analysis was simply replete with difficulties it
would be an impossible undertaking. Fortunately, socio-economic prob-
lems typically offer a multitude of design options, as we saw, for instance
with the DMC problem (Fig. 7.10 in particular). There are usually few laws,
human or scientific, to prohibit the system's structure from being changed.
The policies by which it is managed can be whatever its managers choose to
adopt, or can be persuaded to apply when the results of analysis have been
explained to them. This extensive freedom to redesign a system's structure
and policies makes system dynamics not only a practical proposition but
also a discipline capabl e of providing useful policy guidelines for problems
of great practical and human significance.
Designing good behaviour into socio-economic systems is the essence of
the matter, and the tools and techn iques are simply the servants of that aim.
Even apparently abstract research, such as the model of the decline of the
Mayan empire has, at its roots, a policy orientation, even though the 'man-
agers' concerned are long since dead. The policy emphasis is still present in
models of the dynamics of ecological systems, in which human management
may not be overtly represented. You should now review Chapter 1 and
write down your own ideas in your own words about the fundamental
nature and purpose of system dynamics.
IN WHAT CIRCUMSTANCES?
Chapter 11 described the enormous range of the potential applicability of

system dynamics, but the fact that a methodology can be applied to a
particular problem does not mean that the circumstances are necessarily
propitious to a successful application. We therefore need to consider some
of the factors which might make for success or failure.
The first point we need to establish is that the purpose of any manage-
ment science investigation is to change a situation. This is clearly true of
practical work in a business or in a government department. The object is
usually to support the making of a decision and , whether the study is simply
to confirm whether a given decision is sound or whether it is to help in
choosing the decision, the end result will be to change the system. This
factor of change is still true even when the investigation is into, say, the
processes which led to the collapse of the Maya. The object now becomes to
influence the ways in which scholars have interpreted and understood that
In what circumstances? 349
problem or to help them to organize their insights, and hence generate new
ones, through the construction of an influence diagram, but the net effect
will still be change.
To bring about change to a system, a situation or a point of view is
essentially a political act, not a scientific one . To be sure, it is a good idea if
the technicalities are correctly carried out , which is why we have laid so
much emphasis on checking models and dimensional analysis, but the tech-
nicalities of modelling are subordinate to the political realities.
This line of reasoning suggests that the first consideration in a successful
application is identifying the potential client for the work and discovering
the reasons why that person or group should want to be a client. Unless the
study can be conducted in such a way as to satisfy their demands, its results
are , at best, likely to be ignored.
Use of the term 'client' conjures up images of fee-earning work for
industry or government or of directly sponsored research on a defined
problem. That is, however, too restrictive, and it is helpful to keep the idea
of a client in mind even for abstract academic research on a problem which
excites one 's interest'. The client will be the academic community in, say,
economics . Economists have their own agenda of research, their own un-
derstanding of problems and their own paradigms, and one ignores that at
the risk of wasting one's time on work which will be neglected or derided.
At the same time, it is pointless to pretend to be, for this illustration,
an economist, on the strength of having read a couple of books and audited
a few classes. Truly expert advice is needed, and getting on the same
wavelength as the client is as vital in academic research as in consultancy
practice.
Market identification is rarely easy, but it is essential. The snag is that
clients may not realize that they should be interested in a system dynamics
approach. Overselling is rarely a good idea. Being able to point to successful
applications elsewhere may help, but the fact remains that this stage can be
very difficult. Time, effort and thought must be expended. The author is a
great believer in what he calls 80/20 modelling, which means that 80% of
the time when one's brain and insight are working at full throttle (not the
same thing as 80% of one's working day) should be spent on thinking about
why this problem is important, who thinks it is and why they care about it,
the other 20% being available for clever programming. He has seen rather
too much work which is more like 2/98 modelling.
The second factor is to be able to identify dynamic behaviour in a system,
which is usually not too difficult, and to be able to imagine ways in which it
might be improved, through policy changes and/or structural modifications.
1 The author's research interest at the time of writing is, in collaboration with Professor
Piatelli of the Italian National Research Council, a study of the growth of sea power in the
Classical Greek city states. On the face of it, this is an abstract problem, though one which may
have considerable implications for modern politics and defence str ategy. Who might the clients
be for such work ?
If the system cannot be changed , it is either perfect or paralysed and, in

either case, the study would be a waste of time.
The third step in deciding whether or not to try to apply system dynamics
is to have some idea of the time by which results would have to be available
in order to be able to meet client requirements. Can anything useful be
done within that time and with the resources available?
SYSTEM DYNAMICS IN PRACTICE
System dynamics is a fascinating and challenging approach to policy design.

At a technical level it is easy; at a deeper level it is very difficult because of
what we shall call the 'blank paper problem'. This means that one does not
know in advance what the model should look like. By contrast, the analyst
knows in advance that a linear programming (LP) problem will consist of a
set of linear constraints and a linear objective function. Building an LP
model involves steps which are distinctly non-trivial, but the fundamental
mathematics are well understood and the form of the solution is known in
advance. By analogy with the carver who starts with a piece of wood and
cuts away all the bits which do not look like a duck, the analyst starts with
a problem and cuts away anything which does not look like a linear equa-
tion. In other words, the problem is forced to fit a linear structure, which
mayor may not be legitimate, depending on the problem' . Much the same
is true for most other management science methods; the form of the solu-
tion is known and it is a matter of judgement and knowledge as to how well
that form fits the given problem.
System dynamics, in common with other systems methodologies, such
as the soft systems methodology (Checkland and Scholes, 1990), has no
fixed form and the analyst is faced with a blank piece of paper on which a
model must be created. The essential trick is to realize that one cannot
do this alone : one must work with the client or problem proprietor, and
preferably with a group of three or four people who are involved in the
problem.
The real subtlety is that the clarity of influence diagrams, especially at the
higher levels of the cone, where the conventions of dotted and solid lines
can be dispensed with, is so great that clients will be able to work with the
analyst on model formulation. It is noteworthy that the Mayan problem
involved one system dynamicist and three historians working together.
One of the exciting things which can happen is to have a member of the
client group say 'that diagram is wrong because .. .'. From that basis,
progress can be made with the clients heavily involved in what is going on
2 The author is very well aware that there is a good deal more to LP than that and that there
are integer and non-linear programming methods.
System dynamics in practice 351
rather than them having the feeling that they are detached, or divorced,
from progress and that they will have to leave it to the analyst to come back
later and tell them the answer . That is never going to lead to accepted
results.
Diagram development is best done with a whiteboard and room to walk
around. The group dynamics are such that people will wish to argue, debate,
think and physically stand back and look at what has been done. A good
practical trick is to use the kind of whiteboard that makes a print of what
has been drawn before it is amended too much, or the use of a Polaroid
camera. Using graphical software too soon is bad practice, as it is hard for
more than about two or three people to see what is on the screen, only a
small part of the diagram can be seen at any time, the person controlling the
mouse may come to dominate the thinking and the rest of the group are
likely to lose interest. Software which projects onto a large screen is usually
not helpful because the person running the program cannot be involved in
the discussion and may be perceived as being too much in charge of the
debate, and there is the risk of delays as the diagrams are built and proj-
ected. The essence is to keep everyone involved and active.
An important aspect of a good project is to feed results back as they
arise. Again, one of the attributes of system dynamics techniques is their
rapidity; initial results , whether from a diagram or a model, can often
emerge within days or even hours and showing the client what is happening
is a good way of keeping up his or her interest. This is especially important
in academic work. A given project may lead to a paper based on the
diagram, comment on which may be helpful in revising the diagram and
improving the subsequent model , and it also establishes the project as an
accepted part of the scenery, so to speak. When final results are produced,
they are less of a surprise to the academic community and hence may be
more likely to be taken seriously, rather than being 'butchered with con-
summate skill' (p. 309, footnote 3).
This ability to produce early results quickly has a more profound implica-
tion. It is a truism of analysis that the first model of a problem is always
wrong, usually because the client's need for it has not been properly iden-
tified or the client has been inadequately educated as to what his or her
needs are. Being able to produce an initial model from a few hours or days
of effort means that so little time, money and personal prestige have been
invested that one can abandon the initial model and try again with the
benefit of the improved understanding. Any method which requires person-
months of effort before any results emerge will produce models which are,
psychologically and politically, almost impossible to amend, even if they
profoundly miss the point of the problem.
Sufficient has been said in earlier chapters about the importance of not
making the simulation model over-complicated, and we have also explained
that there is nothing in most of the software packages to prevent one from
modelling exactly as much complexity as is needed. In other words, the

problem can be modelled just as it needs to be represented and there is no
need whatever to force the problem to fit a given format.
The fact that even quite complicated models will run in a matter of
seconds on a PC means that client involvement can be maintained and
exploited through the medium of study periods. These are intensive work-
ing sessions of about one or two days, based on groups of three or four
client staff, or academic colleagues, each assisted by someone familiar with
the model and adept at running it. The results which emerge are perceived
by the client as belonging to the client, and this sense of ownership of the
answer to the problem of which they are also the proprietors does wonders
for acceptance of the value of a study . Until one has taken part in a study
period one will not realize the benefits they produce'. Study periods with
academic colleagues from other backgrounds are a valuable part of interdis-
ciplinary academic research.
Study periods are particularly valuable as a precursor to optimization. As
we have seen, optimization of models is a very powerful approach, but, as
with all strong medicine, it needs to be taken cautiously; it is vital to
optimize the right model and to optimize the right aspects of the problem.
Careful attention to the discussions in the study period will usually throw a
good deal of light on what the problem proprietors consider to be important
in the system's behaviour. That leads to more acceptable objective func-
tions. It sometimes happens that there are two or more objective functions
which represent the aims of competing interest groups within a problem.
The fact that a thorough optimization search can be done within a few
seconds' allows one to run study periods devoted to optimization of candi-
date, or competing, objective functions with the aim of finding policy op-
tions which satisfy everyone.
These are all positive aspects of practical work. The negative side is that
there is an acute danger that the model will take on a life of its own. Indeed,
the phrase 'in reality the model shows ...' sometimes occurs in published
work. This is nonsense. The model is not reality; it is an artefact and a
simplification and it is essential not to lose sight of that.
Presentation of results is an important aspect of practical work and has
three aspects. The first is the influence diagram. It is usually disastrous to
show the complete diagram as a single illustration in a report or paper.
Building the diagram up through several stages is often helpful (see, for
example, the police effectiveness case study in Chapter 11 and Coyle
(1992)). Using different diagrams from the various levels of the cone can be
useful. Do not assume that the diagram is self explanatory; a picture does
3 The author once led a study period at which a rather tasty chicken dish was served at lunch
on the first day, but not on the second. When the part icipants complet ed their appraisal forms,
the only improvement suggested was that chicken should be served on both days. There spoke
some happy customers!
4 This depends heavily on the capabilit ies of different software packages .
Assessing the quality of a system dynamics model 353
not save 1000 words - it may well require them. Secondly, presentation of
the equations requires some thought. A full listing of a model is not always
helpful , and documented printout of the most significant equations, sup-
ported by a few sentences of explanation, may be more appropriate. It
convinces clients of the quality of the work, without overwhelming
them with detail. Finally, the recommendations for policy or structural
change need to be handled with care. It is often a good idea to present the
results in such a way as to show clients two or three things about the system
which they already knew or suspected as well as the novel insights and
recommendations which have emerged.
The essence of good system dynamics practice is to bear in mind that the
model represents a policy problem for a client and it is the policy which is
the object of study, not the model.
ASSESSING THE QUALITY OF A SYSTEM DYNAMICS MODEL
A critical aspect of good system dynamics practice, whether it be in consul-

tancy or in academic research, is to ask a colleague to review one's model.
This exposes the model to, hopefully constructive, criticism and is an essen-
tial step in quality assurance. The review should be done more than once
during a project. The following are some of the criteria which can be
applied.
Purpose: Has a clear purpose been defined? Is it a worthwhile aim, likely to

help in the solution of a real problem?
Suitability: Does the model meet its purpose? This is rarely clear-cut. It
depends on some of the following criteria, and the question ought really
to be answered at the end of the review. Posing it at an early stage
is a useful way of creating the right frame of mind for the remaining
questions.
Basis: Is the model soundly based on accepted theory of the processes in the
problem, if there is any, or does it contain sweeping generalizations? The
question ought to be stated in this provocative form if later, and probably
much more hostile, questioning by the client is to be avoided . Failing
accepted theory, what assumptions and simplifications have been overtly
stated, preferably in writing, and what others have been made implicitly.
The model builder often does not realize what has been assumed.
Credibility: Does the sponsor attach value to its results? If the results are
unexpected or provocative, how can we ensure that they are seen as
interesting and valuable, as opposed to controversial and unacceptable?
Creativity: Does it provide new insights into problems and suggest new
solutions? A model which did no more than confirm accepted wisdom
would verge on the sycophantic and would be a most unusual event in
system dynamics work.
Simplicity: Has the real world been simplified to just the right extent ? This
is rarely a technical matter and often has much more to do with showing
a particular interest group that 'their' part of the problem has been
included. The difficulty is that it can lead to intractable models. The real
subtlety is that the model must be simple enough to be tractable while
ranging widely enough to contain the possible solutions to the problem
and to ensure creativity. The reader might like to review the case studies
in Chapter 11 to see how well they meet this test, bearing in mind that
case studies in management education are never intended to show exam-
ples of good or bad practice; they are intended to make the student think.
Redundancy: Can parts of the model be removed without affecting its
behaviour" This does not necessarily mean that the model has not been
simplified enough . It often means that aspects of the problem which
everyone knew to be important are not significant. In the psychology and
politics of studies, this may cause some part s of the client group to heave
a sigh of relief when they can prove that the problem was not their fault.
On the other hand it may upset them deeply to discover that they are not
as important as they thought they were. Careful handling will be called
for, and it is as well to discover this kind of result early enough for
precautions to be taken .
Transparency : Are the model's assumptions, simplifications and equations
well understood by analyst and sponsor? If not, the analyst is working
blind and the sponsor has to take the results on trust. Study periods
should have ensured that the model and its results are the shared prop-
erty of analyst and client.
Flexibility: Can the model deal with new inputs and parameter values?
Generality: Can the model represent other problems, perhaps when param-
eter values are changed ? It is often poor practice to try to make a model
too general. The value of a model depends fundamentally on how well it
is suited to its purpose and a general model may be trying to satisfy too
many purposes . System dynamics models are typically quick and cheap to
develop and it is generally better to develop a new model for a new
purpose.
Sensitivity: How severely does it react to changes to input data, parameters
or initial conditions? Testing a model under extreme conditions is an
essential step to verifying that it is technically correct.
Soundness: Does it satisfy technical criteria such as dimensional consist-
ency, accuracy of mass balances, correct use of DT and legitimate use of
rate-dependent rates? In other words, within the framework of the as-
sumptions and simplifications, does it do the same things as the real
system and for the same reasons ?
5 The COSMOS optim ization software includes a simplification technique designed to show
which parts of a model can be switched off without degrading the final value of the objective
function by more than X%.
Teaching system dynamics 355
Productivity: Is it easy to use for experimentation and optimization?

Cost: Is it expensive or slow to develop?
Promotion : Are the claims made for the model's results reasonable or
exaggerated?
A thorough review of one's work can be a deeply uncomfortable process. It

is rather akin to having one's cherished research butchered by a journal
referee who shelters behind his or her anonymity and abuses it to show how
clever the referee is and how he or she would have done the work. Nonethe-
less, the experience must be endured partly because it is an inevitable
component of true professionalism, but also because it may well avoid a lot
more grief later at the hands of the client. This is especially true when the
work is abstract research and the client is some other part of the academic
community.
TEACHING SYSTEM DYNAMICS
System dynamics is difficult to teach ; at any rate the author has found it to
be so, despite many years of trying. To be sure, the basic techniques of
influence diagrams and simulation are easily understood, and most students
have no difficulty in following models developed during class periods, run-
ning them and doing policy experiments and optimization. The difficulty
arises when they subsequently try to tackle a problem of their own, without
guidance. Forrester has always argued that to become a good system
dynamicist one must spend a period of apprenticeship with an experienced
practitioner but , while this works well for the student who can spend a year
or more on a higher degree , it is a counsel of perfection, as most students
cannot do that.
One of the purposes of textbooks such as this, and especially the prob-
lems and cases it contains, is to provide students with synthetic experience
and reference material to help them over this hurdle. The essence of the
teaching problem is, however, the same as the difficulty in practising system
dynamics, discussed earlier, which is that one has to start with a blank piece
of paper, and many people find that to be very hard. The only thing the
instructor can do is to be aware of the problem and to try to frame the
teaching to deal with it. The purpose of this section is to try to pass on some
experience and suggestions to those who may wish to embark on their own
teaching of system dynamics.
Although helping students to develop a deep understanding of system
dynamics is difficult, it is easy to construct a course of lectures and practice
to give at least initial competence in the discipline. What follows is an
outline of a one-week intensive course intended to start at about 11 a.m. on
Monday and finish at lunchtime on Friday, a common pattern for short
courses . Students are expected to undertake evening reading.
Figure 12.1 shows a timetable based on periods of 50 minutes, as students

from industry are not used to sitting in a classroom and frequent breaks to
stretch their legs and their brains are essential.
The course is geared to modelling practice so that, at the end of the week,
the student has built, debugged, analysed and optimized two models. The
software used is irrelevant, as the focus is on the principles and fundamen-
tals of system dynamics, not on a particular syntax and graphical front end.
The production model is the Domestic Manufacturing Company's problem,
studied in earlier chapters. Since most of the students come from a defence
background, the second model is the simple combat case study. Obviously,
this is easily varied to suit different student groups. Readings can be as-
signed from chapters of this book or from the suggestions made below for
further reading in system dynamics.
The author has experimented with many approaches to equation writing
with students of various backgrounds in different universities. The ap-
proach he now uses may seem old-fashioned, but it does work", The method
is to write the equations on a whiteboard, with the students copying. Three
boards are used, one for equations, another for definitions and dimen-
sions and the third for explanations of, say, the smoothing equation. The
author's handwriting is less than perfect and he may make mistakes (even
deliberately). Students are likely to make copying errors and will probably
make more when they type the model into the computer.
The end result is that the model is practically guaranteed not to work, the
point being that students will learn far more about system dynamics by
debugging the model than they ever will by being given a perfect version to
start with (an instructor solution is issued afterwards, of course). To save
time, the author once issued each student with a personal version of a
model containing deliberate mistakes, and the reaction was universally that
the students felt that they would never make such silly errors. He now
leaves them to find for themselves just how silly are the mistakes they can
make. Dimensional analysis is a key part of this process. Although COS-
MIC has a graphical model-building front end it is not used in the instruc-
tion as the emphasis is on modelling principles, not on a particular package.
This approach works well with class sizes of up to 30, but would be
unlikely to be successful with larger groups.
The timetable is easily adapted to give a course of about two periods per
week over a term or semester. Parts of the material can be used to give a
more condensed treatment of system dynamics within a management
science syllabus, though teaching system dynamics as part of a course in
simulation techniques is missing the point; simulation is only a tool
6 At the end of the course, the students, who are usually experienced modellers , complete a
'happiness sheet' in which they rate each topic according to the technical level at which it was
presented (too low, about right, too high) and the time spent (too long, about right, too short).
The course is normally given a practically perfect score.
Time Monday Tuesday Wednesday Thursday Friday
08 :50 Principles of Modelling practical Development of a Advanced aspects of

09:40 simulation combat model SO modelling II
09:50 System dynamics case

10:40 study II: optimization
of defence expenditure
11:10 Introduction to system Formulation of Optimization in SO Guest speaker: SO Applicat ions of SO

12:00 dynamics production model in practice
12 :10 Introduction to Optimizat ion of Rev iew of SO software

13 :00 influence diagrams production model and course discussion
14:10 Develop IDs for Load, debug and Advanced aspects Running of the Optional modelling
15:00 production model run production of SO modelling I combat model practical
and combat model model
15:10
16:00
16:10 Influence diagramming Policy design in System dynam ics Optimization of the
17:00 problems production model case study I: combat model
qualitative analysis
17:10 Influence diagram Experiments with Modelling practical Modelling pract ical
18:00 work production model
Fig. 12.1 Outline program for a system dynamics course.

by which the underlying thought patterns of system dynamics are

implemented.
Apart from teaching system dynamics modelling per se, the format of Fig.
12.1 is also used to teach information system modelling to students on a
masters degree in information systems. The DMC model is still used, but
with the emphasis on the information structure, its interaction with policy,
and the consequences these have for system behaviour. Only one model is
built, and other sessions are used to consider applications of system dynam-
ics to other problems. Such a course occupies about 25 lectures.
Students on Master of Business Administration courses may take system
dynamics as a course in a management science or strategic management
option, essentially using the timetable of Fig. 12.1, though with different
models as the case studies. For students not doing such options a course of
about eight lectures and cases can be structured around influence diagrams
under the title of 'Organizing Complexity'. The theme is that management
problems are inherently complex and dynamic and that understanding the
mechanisms of dynamic behaviour is an important aspect of strategic think-
ing. The diagrams used are for business firms, socio-economic problems
and industrial structures. The last couple of lectures are used to show
what a simulation model can do and the effects of policy and structural
optimization .
FURTHER READING
Any book reflects its author's prejudices and experience , and a subject,
especially one as subtle as system dynamics, can only be deeply understood,
as opposed to technically grasped , by reading of other work and the absorp-
tion of other views and prejudices. This section outlines some suggestions,
in chronological order of publication.
The author first read Forrester's masterpiece, Industrial Dynamics
(Forrester, 1961) about 30 years ago and he still re-reads it from time to
time, always to his profit. Since system dynamics was established, numerous
authors have written texts on the subject.
Richardson and Pugh's Introduction to System Dynamics Modeling
with DYNAMO (Richardson and Pugh, 1981) is a good example of this
literature. It provides a number of examples of simple models and is,
to some extent, a supplement to the DYNAMO User Manual (Pugh
did a great service to system dynamics by developing the original
software).
Roberts et al. (1983) cover the techniques of using the DYNAMO soft-
ware and offer a wide range of examples, most of the models being rather
simple.
The common themes to both these books are that only DYNAMO is
mentioned and that nearly all of the models they contain are completely
Finally, on a personal note 359
closed in that the dynamics arise solely from the structure of the model , its
parameter values and initial conditions, and not from any exogenous influ-
ences. This might be termed the American school of system dynamics , as all
the authors were for a long time associated with Forrester's group at MIT.
The view adopted in this book is that both closed and driven systems exist
(the combat model and DMC are, respectively, examples) and that one
needs to know how to deal with both.
Both books are now some 12 years old and a more recent offering is
Wolstenholme's System Enquiry (Wolstenholme, 1990). The examples it
contains are from business, coal mining and defence analysis.
A useful comparison of system dynamics and event-based simulation is
provided by Pidd (1992).
Finally, the System Dynamics Society's journal, System Dynamics Review
(see p. 309, footnote 4) provides an invaluable picture of current work in
the field.
DEFINITIONS OF SYSTEM DYNAMICS
In Chapter 2 several definitions of system dynamics were put forward . You

have , we hope , pondered these as you have worked and you should now
understand this fascinating and wide-ranging subject fairly thoroughly. Per-
haps, as a culminating task, all that knowledge and study can be put to use
to formulate your own definition of the subject. The author would be
delighted to hear from you with your suggestions .
FINALLY, ON A PERSONAL NOTE
I first encountered system dynamics in 1967 at the feet of a master, Profes-

sor Jay Forrester. I had read his book before that, but the subject really
came alive at his hands and I owe him a debt.
Since then, most of my working life and, indeed, many of my waking
hours have been spent doing and thinking about system dynamics. Apart
from providing me with an academic's typically modest living it has given
me an immense amount of intellectual and practical enjoyment. The intel-
lectual part came from dealing with some challenging problems and work-
ing with many excellent students, many of whom subsequently became
respected colleagues and good friends . Practical enjoyment came from
tackling some demanding consultancy assignments and feeling that I was
doing something useful in the real world.
A great personal pleasure is that my son, Jonathan, who wrote part of
Chapter 11, also enjoys system dynamics and is very good at it.
This almost reads like a premature obituary, but I hope to continue in the
same vein for many years yet, God willing. In fact, the labour of writing this
book has been a rewarding chance to draw together thoughts which have
been developing during the 15 years since my first book on system dynamics
(Coyle , 1977).
My final word is that I hope that the hard work you have put in to
studying this book will help you to enjoy system dynamics as much as I do .
REFERENCES
Checkland, P. and Scholes, J. (1990) Soft Systems Methodology in Action, John

Wiley & Sons, Chichester.
Coyle , RG. (1977) Management System Dynamics, John Wiley & Sons, Chichester.
Coyle, RG. (1992) The optimisation of defence expenditure, Europ ean Journal of
Operational Research, 56, 304-18.
Forrester, J.W . (1961) Indu strial Dynamics, MIT Press.
Pidd , M. (1992) Computer Simulation in Management Science, 3rd edn, John Wiley
& Sons , Chichester.
Richardson, G.P. and Pugh , A.L. (1981) Introduction to System Dynamics Modeling
with DYNAMO, MIT Press (reissued by Productivity Press , undated).
Roberts, N. et al. (1983) Introduction to Computer Simulation: A System Dynamics
Modeling Approach, Addison-Wesley, Reading MA.
Wolstenholme, E.F. (1990) System Enquiry: A System Dynamic s Modelling Ap-
proach , John Wiley & Sons , Chichester.
APPENDIX A
System dynamics software

packages
INTRODUCTION
System dynamics models can be written, with some effort , using spread-
sheet packages or general-purpose languages such as FORTRAN but there
are several commercially-available software products specifically for system
dynamics modelling. The author has, however , seen some remarkably good
system dynamics models written, laboriously, in FORTRAN and some
quite unbelievably bad models written, very easily, using some of the pack-
ages described here. In general, good quality system dynamics is more to do
with intelligent and diligently careful use of the package one has available ,
and less to do with having any particular one, though the packages differ
substantially in the tools they provide.
This book was written using those parts of COSMIC which are consistent
with traditional formats for system dynamics models, but COSMIC has
many other features, as have the other packages . In an attempt to be fully
comprehensive, this Appendix provides a brief summary of the available
packag es.
The Appendix falls into thr ee parts . The first is an outline of the historical
pedigree of the different packag es, indicating the degre es of consistency
and difference between them.
The second part is composed of summaries of the capabilities of the
several packages , provided by the companies concerned. The author has
eliminated qualitative assertions , such as that Software X is 'easy to use' so
as to provide a factual summary, though the author accepts no responsibil-
ity for the accuracy of the information.
Finally, the author offers his own suggestions, based on many years
experience of doing system dynamics, of some factors to be considered
when selecting a software package.
HISTORICAL PEDIGREE
To put the individual packages into context a brief discussion of their

historical pedigree may be useful. Figure A.I shows the lines of evolution.
362 Appendix A : System dynamics software packages
Optimisation DYNAMO
methods ,/ (1960)
derived from .,/ ,/
numer ical -
mathematics DYSMAP
I '\ ,/ (1970)
I '\~ .//
I COSMIC
I and
COSMOS STELLA/
I (1984) iTHINK
I (19 \ \
I
~ \
VENSIM
(1990) \
185)
POWERSIM
1 ( 1992)
+
Thepresent day
Fig. A.I The evolutionary lines of system dynamics software.
The solid vertical lines show that the package is still available, but does not
show the several stages of development that each has gone through. For
instance, the DYNAMO which was available in 1960 bears little relation-
ship to what is now produced, DYNAMO, DYSMAP2 and COSMIC,
though originally developed for mainframe computers, all now exist in PC
versions.
The dotted lines indicate that one package is very similar in operation
and syntax to its 'predecessor' , though not, of course, plagiarised from it.
For instance, ithink and Powersim are both based on Forrester's flow
diagrams . It should be understood, however , that although Package Y may
be 'descended from' Package X it is very likely to contain features which
Package X did not have and, of course , since the evolutionary split X itself
may have acquired facilities which Y does not possess. Indeed, in some
cases, facilities which were new to the 'descendant' have since appeared in
the 'predecessor'.
In short, the only reliable guide to what a given package does , and does
not, support is careful reading of the relevant manual and , where possible,
close comparison with other manuals . Even better is to use the packages or,
failing that, to talk to users.
The three groups of packages are very different from each other. The
DYNAMO/DYSMAP/COSMIC evolution is derived from the concept of
writing equations using a text editor, though COSMIC also has a graphical
modelling support system. The STELLAIPowersim line is geared to the use
of symbolic icons for drawing level/rate diagrams, from which some of the
equations are written automatically. Vensim is rather different from both
these groups, being derived from the text editor approach but also support-
ing icon-based modelling.
SUPPLIER SUMMARIES OF THE PACKAGES
Introduction
The following paragraphs were provided by the software suppliers. The
descriptions are in alphabetical order of the product name. All the product
names and italicized features are trade marks of the respective suppliers
and all trade marks are acknowledged.
To emphasise that this is what the suppliers chose to say about their own
products, in response to the author's invitation, the text is shown in quota-
tion marks.
The COSMIC and COSMOS software

'COSMIC and COSMOS are an integrated environment for the develop-
ment , analysis and optimisation of system dynamics models. The software
suite is based on a Workbench concept and can be run on any Pc. It can run
under DOS or WINDOWS.
Model building is supported by a graphical development environment in
which the user builds Influence Diagrams on the screen, creating and defin-
ing new variables as required. The Influence Diagram data are stored and
can be used for model building. Level equations are written automatically;
rates and auxiliaries are created on the screen by the modeller using a
mouse and the appropriate part of the Influence Diagram. As equations are
created, a fully-documented model is stored.
COSMIC has a total of 44 built-in functions and interactive guidance is
provided on their correct use. TABLE functions are created by drawing
graphs on the screen. The graphical environment also provides facilities for
editing diagrams and equations while maintaining the consistency between
them, for tracing paths and loops in models, and for browsing equations and
the associated Influence Diagrams.
Alternatively, the modeller uses the text editor to type in the equations in
the traditional manner. The modeller can change from the graphical work-
bench to the text editor at any stage in model development.
Debugging and running of the model is supported by extensive diagnostic
software including dimensional analysis, detection of feedback loops, log-
ging of simulation runs and statistics of function usage. Output can be
produced as tables or graphs , the latter being available on the screen (with
hard copy options), on the printer, or on a flat-bed plotter with user control
of line pattern and pen colour. The model can be saved as a FORTRAN
program.
COSMOS supports four methods of optimization. Direct optimization

uses a hill-climbing algorithm to design policy options or to fit a model to
parametric or historical data. Base vector optimization tests the sensitivity
of the model to subsets of parameters which are chosen from candidate sets
provided by the user. Simplification is used to test which parts of a model
have the most effect on behaviour. The Planning Horizon method can be
used to study the ability of a model of a system to adapt itself to changing
events .
For more information contact The COSMIC Holding Co, 8 Cleycourt
Road, Shrivenham, Swindon SN6 8BN, UK. Telephone +44 (0)1793
782817, or use the order form at the end of this book.'
The DYNAMO software

'DYNAMO allows the user to create large models (via a text editor) using
arrays, macros (either built-in, such as DELAY, or definable) and functions
(either built-in, such as TABLE, or definable in C). Models can be docu-
mented to collate all uses of a variable. Compilation errors are shown via an
integrated editor. Simulation results can be stored and DYNAMO can
integrate with spreadsheets for data input or output.
To support ownership of models, the DYNAMO RepW scripting lan-
guage supports templates, FOR loops, macro actions to permit 'what if', or
a gaming system under Windows 3.1. This system permits up to 4 views
(Reports, Plots or Flowcharts) on a single screen. Reports analyse or
explain simulation results and prompt the user to change parameters with
on-line help . Plots displays curves. Flowcharts or other diagrams can be
imported for display.
For further information contact: Pugh-Roberts Associates, 41 Linskey
Way, Cambridge MA 02142, USA. Telephone +1 717864 8880, Fax 864-
8884.'
The D YSMAP2 software

'DYSMAPZ accepts models from a source file. Equations may be written in
any order and variable names may be up to 32 characters. Syntax errors are
reported to the screen and to an optional file and run-time errors are
reported. The software supports dimensional analysis and includes the
usual logical, time-related and delay functions. Delays of orders 1, 2 and 3
and a pipeline delay are provided.
After execution of a model the user is placed in an Interactive Command
Environment (ICE) with all variables saved for subsequent plotting. Com-
parative and scatter plots can be performed. Zooming on the time axis is
permitted.
Up to 10 runs can be performed in anyone session and new runs can be
launched interactively from the ICE. These usually follow the Change or
Insert commands which allow the user to respecify any equation, table or
constant in the model , or introduce a new variable, table or constant.
Input and output includes reading table data direct from file and output
of numerical data in tabulated or ASCII free-format.
For further information contact Salford Software Ltd , Adelphi House,
Adelphi St, Salford M3 6EN , attention Dr Olga Vapenikova. Telephone:
+44 (0)161 8342454, Fax: + 44 (0)161 8342148.'
The ithink software!

'The ithink software comes in two versions: Core and Authoring. The Core
version is designed to enable people to construct system dynamics models
by piecing together stocks, flows, convertors and connectors on an elec-
tronic tableau. In addition to the standard Reservoir, the software includes
three additional, specialized, stock types: Conveyors, Queues and Ovens.
Other facilities include automatic cycle-time calculations and statistics,
drill-down sub-model capability, high-level mapping capability, unit-con-
version within flows and space-compression objects (for managing visual
complexity) . The package supports multi-variable sensitivity analysis and
run-by-sector capability. The documentation comprises: An Introduction to
Systems Thinking, The Process Impro vement Module , The Business Appli-
cations Guide and The Technical Reference Manual.
The Authoring version of the software enables people to create run-time
accessible, stand-alone Learning Environments, or Management Flight
Simulators. No knowledg e of programming is required to create these high-
level interfaces, which can include Quick/Time movies, slider inputs , graphi-
cal function editors, on-condition message posting and numeric readouts.
Authors can regulate how much access to allow end-users to the underlying
model structure which they have created.
For more information contact High Performance Systems Customer
Service at +16036439636 (9 am-5 pm Eastern Standard Time, USA), Fax
+ 1 603 643 9502.'
The Powersim software

'Powersim is a Windows-based graphical system dynamics modelling pack-
age. It supports both flow diagrams and causal loop diagrams. The underly-
ing equations can be entered when the model's structure is defined.
Equations are created with the aid of a library of built-in functions and a
visual editor for drawing graph functions.
Simulation results are presented as animation, numbers or graphs.
Powersim also offers dynamic objects which display selected variables as
time graphs , scatter graphs , time tables , bars and numbers.
I The ithink software also exists as STELLA II, which is marketed to academic institutions.
Powersim supports the creation of games and "management flight simu-

lators". Parameters can be changed using slider buttons . Push buttons,
radio buttons and check boxes can be added to the gaming interface and
data can be imported and exported using additional modules.
Powersim supports multidimensional arrays. The results of a simulation
can be saved and compared with other results. Sensitivity analysis can be
performed by setting up multiple simulations with varied inputs. Separate
models can be connected and simulations run in parallel.
For more information contact ModellData AS, Zachariasbryggen, PO
Box 642, N-5001 Bergen, Norway. Telephone +47 55 31 3238, Fax +4755
31 31 36.'
The Vensim software

'Vensim is an integrated environment for the development, analysis and
application of system dynamics models. It supports causal loop diagrams
and stock and flow diagrams. Information from the diagrams is stored to
allow the creation of model equations. It provides Document, Loops, Out-
line and Tree tools to generate different representations of model structure
and includes a dimensional analysis facility.
Any number of named runs or scenarios can be created and there are Bar
Graph, Gantt Chart , Graph, Stats, Strip Graph and Table tools for the
analysis of data and simulation results. Vensim's tools support Causal Trac-
ing, facilitating the review of simulation results and the tracing of underly-
ing causes.
Reality Check will test families of models under a variety of assumptions.
Optimization and data handling tools support calibration of a model
against historical data and the selection of policy interventions.
The scripted application language can be used to develop learning envi-
ronments, game and decision support systems with access to the analysis
tools. The software supports subscripts, special variable types and user
defined macros. An interface can be used to create special purpose
functions.
For more information contact Ventana Systems Inc., 60 Jacob Gates
Road, Harvard, MA 01451, USA, Telephone + 15084563069, Fax + 1508
456 9063, Email vensim-world.std.com..
FACTORS IN SELECTING A PACKAGE
Introduction
The newcomer to system dynamics will need software support and, as
we have seen, there is a fairly wide choice. Even the established practitioner
of system dynamics can find it useful to have more than one package
available. The price differences between packages are not large and, in any
case, the price is trivial relative to the power they confer. How is one to
choose?
In this section, we shall attempt a summary of the features typically found
in system dynamics software , working from the available information, but
reasoning from first principles. Where it is necessary to mention particular
packages by name , alphabetical order is used.
The views expressed are very much the author's opinions. They are based
on much experience of system dynamics, but others might disagree with his
views. Recall also, that all the packages are undergoing continual develop-
ment so, if software X is criticized for lacking feature Z it may well be that
Z will appear in X in due course. If software X is mentioned as having
facility Z, it is simply to illustrate the argument and does not necessarily
imply that other packages do not have an equivalent of Z.
In general, a system dynamics software package needs to be assessed
according to
• its basis in fundamental system dynamics theory;
• the ease with which it can be used;
• the support it gives to model building;
• the extent to which models can be documented and explained to a
sponsor;
• the facilities it has for debugging a model ;
• the ease of making experiments and producing output;
• the scope of its facilities for policy design.
Speed of running a model is nice to have but the critical feature is the
support of thought, not the crunching of numbers.
Fundamental theory
All the packages except one are based on the fundamental Level, Rate and
Auxiliary variables which were discussed in Chapter 4, especially Figure
4.2. The exception is Vensim which only recognizes Level and Auxiliary
variables in its text editor, though allowing icons for rates in its graphical
tool.
DYNAMO, DYSMAP2 and COSMIC identify different types of
variable by initial letters in the equations, such as L, R and A. ithink
and Powersim identify them from the icon used to draw that diagram; using
a box symbol will create a level variable. Vensim identifies variables by
the form of the equation defining the variable and optionally allows the
use of icons and built-in typographical conventions to highlight variable
names; the name for a constant, for instance, will be displayed in UPPER
CASE.
368 Appendix A: System dynamics software packages
Ease of use
DYNAMO and DYSMAP2 depend on models being typed into a file from
a text editor. COSMIC, ithink, Powersim, and Vensim have Graphic User
Interfaces (GUI) which allow one to draw diagrams on the screen as a
precursor to equation writing. In all cases, a mouse is used to link the
appropriate symbols into diagrams . Powersim and Vensim also provide
'causal loop diagrams' (see page 43).
The advantage of drawing a diagram on the screen is considerable; the
assumed influence of one variable on another is abundantly clear, as it is in
an Influence Diagram. In both cases, however, the form of the influence is
not clear. The disadvantage with ithink and Powersim is that the diagram
must correspond exactly and in every detail with the equations which are to
be created, because, in those packages, the GUI is the portal to the creation
of equations. If the model is to become larger , so must the diagram . If the
model is to have very complicated logic, then the diagram must have a very
complicated structure, though Powersim's array feature can reduce diagram
complexity and ithink's space compression object allows complexity to be
shown or hidden as required.
COSMIC and Vensim, on the other hand , while supporting GUIs
(though the two are very different), allow one also to enter equations using
a text editor. This allows the user to take advantage of the distinction
between Level 3 and Level 4 in the cone of Influence Diagrams in Figure
2.16. Level 3 is drawn to show the main features of the problem. Level 4
need not necessarily be drawn, as it is implicit in the equations, and the
connection between the two levels is carried in the analyst's mind or can be
inferred from the code if needed.
All the GUIs will, to some extent, write some of the equations automati-
cally, and it is sometimes believed that these packages 'write the model
automatically'. This, however, is not the case. The most that happens is that
the Level equations are written automatically, but the Rate and Auxiliary
equation still have to be created by the user. Creating initial conditions for
levels must also be done by the user. This can be an important part of
modelling and it is worth checking how much flexibility a given package
offers.
In the author's opinion, heavy reliance on a graphical user interface
can sometimes lead the modeller into very bad habits . It is simply too
tempting to rush ahead with putting equations together and then looking
at graphs on the screen rather than carefully checking that the equations
are correct and do the same things as the real system and for the same
reasons.
In summary , all the packages are easy to use, once one has mastered the
appropriate procedures, though the beauty of a GUI is, perhaps, very much
in the eye of the beholder. What to one person might seem elegant and
powerful may be over-complicated and confusing to another.
Support to modelling
A package will support modelling by the extent and diversity of its built-in
functions. For example, delays are a common and significant feature of
dynamic problems so a system dynamics package which did not support
functions for modelling delays would be next to useless. In fact, all the
packages do have greater or lesser numbers of various forms of delay
function. COSMIC has 7 versions of delay functions; ithink has special
functions called 'Conveyors, Queues and Ovens ', which produce discrete,
event-orientated delay processes, though a standard DELAY3 would have
to be written as a sub-model.
In general, however, all the packages are fairly comprehensively
equipped with standard functions though, before selecting a package , the
reader should check exactly what functions a given package does and does
not support.
Model documentation
The process of describing what a model does, what it contains and how it
works is called 'documentation'. It is vitally important to the production of
good models as it enables both the creator of the model and the customers
for it (whether they be fee-paying clients or the academic community) to
understand the model. Without understanding, the model will not be ac-
cepted as a basis for management action or scholarly progress and will
therefore be a waste of effort. The quality of a software package is, there-
fore, more profoundly affected by the extent to which this vital process is
supported than by the elegance and sophistication of its GUI.
The original DYNAMO concept provided for 'maps' of the connections
between all the variables and for documentation of the equations. This
concept was followed almost exactly in COSMIC (DYSMAP2 and Vensim
achieve a corresponding effect in a different way) and has been illustrated
in this book whenever individual equations have been discussed. All the
variables in an equation are defined and any constants used in the equation
are searched for and shown, together with their definitions, even though the
constant may be scores or hundreds of lines away from the equation being
documented. See, for example , page 116. The other three packages have
similar features.
Facilities for debugging

Debugging a model means the process of clearing it of errors of syntax and
logic.
All the packages provide syntax error messages if, say, a closing paren-
thesis has been omitted. All of them ensure that the model will not run if
variables are used without having been created. In short, all the packages
do a thorough job of checking for errors of syntax.
The extent of displaying the error messages is rather more variable

between packages. The DYNAMO concept, reflected in COSMIC and
DYSMAP2, is to provide a printout of all the error messages, which, com-
bined with the model map, enables one to detect related errors in equations
which may be hundreds of lines apart in the model. The other packages
depend more heavily on showing error messages on the screen, which can
make it difficult to see where the true source of an error lies, apart from
such obvious cases as missing parentheses. Printout of error messages is,
however, also available.
Checking a model for errors of logic is more difficult and is a function
both of the steps the modeller has taken to build checks into her model and
the features the software supports.
Built-in checks have been illustrated in several places in this book . By far
the most important is to include mass-balance checks for the physical flows
in the model.
Software support is rather variable. The author's experience, based on
hundreds of models, is that it is completely impossible to debug a model
simply by looking at a few graphs on a screen. Recall the discussion of the
simple combat model in Chapter 6, page 178, in which the model's dynamics
looked right even though the model was still incorrect. Even if one graphed
all the variables, only about six can be shown on the screen at anyone time
and the precision of the graphs is too low for fine details to be seen. In short,
the only way to be sure that a model is doing the same things as the real
system and for the same reasons is the careful examination of printed tables
of the values of variables against time. This was the original DYNAMO
concept and it is unquestionably the correct one. All the other packages
support printed output.
Apart from printing out values of model variables, the most powerful
check is dimensional analysis of model equations. The importance of
dimensional analysis cannot be overstated. The author has worked on
some hundreds of system dynamics models, experience which has proved
to him that he is very good at making mistakes and that software for
dimensional analysis saves any amount of worry over the quality of his
models.
Ease of making experiments

The concept employed in the original DYNAMO was to allow the user to
define a set of experiments by providing new values of constants or tables
followed by a run command . This procedure could be repeated a certain
number of times (up to 50 in COSMIC) to provide a set of experiments
which would be executed in one run of the model. This approach was
essential with Mainframe computers, though it is still essential with PCs,
provided one has a suitable printer. COSMIC extends its value by providing
a run logging option which prints out exactly the experimental conditions
Summary 371
defined by the successive changes of constants and tables. Although there is
some risk of becoming buried under paper, the batch experiment, followed
by careful study of the tabulated values of model variables under different
run conditions is an invaluable way to debug a model, and experiment with
it in the fashion described in Chapter 7.
All the software packages also provide for interactive experiments in
which one changes parameters as required and runs the model again. The
packages vary a little in what they offer. Powersim, for instance, allows
parameters to be changed during a simulation run.
All the packages provide some means of saving the results of model
experiments. For example , COSMIC allows the parameter values to be
printed and the screen graph to be copied to a laser or colour printer.
In summary, all the packages make it fairly easy to perform model
experiments though , as remarked above, it is the author's experience that
heavy reliance on graphical interaction can result in serious errors. To
develop quality models it is vital to review equations carefully and get
printed output documenting the model and giving detailed simulation
results.
Policy design facilities

All the packages make it easy to perform experiments, none of them
hamper the analyst's intuition and all can be said to possess good policy
design facilities.
We have, however, attempted to demonstrate that optimization is poten-
tially a very powerful policy design tool.
COSMOS allows the user to choose the objective function so that it
can be maximized or minimized as the nature of the problem requires,
typically finding the optimum of a large model in about 30 iterations, taking
some 15 seconds on a 486DX2 PC running at 33 Megahertz. There are also
facilities for three additional forms of optimization, beyond the simple
hill-climbing, or direct optimization, which was described in the earlier
chapters.
SUMMARY
All of the packages are very impressive and required much effort to de-
velop, but, before selecting one, try to probe into what facilities it actually
has. A few criteria are given below.
The essential distinction is between the COSMIC/DYSMAP21
DYNAMO grouping, which offer the traditional ability to produce
extensive output on paper for debugging and documentation, as well as
differing degrees of graphical capability, and the ithinklPowersim/
Vensim concepts which are rather more heavily geared to the use of
screen graphics for building and running models, though they can pro-
duce differing degrees of paper-based output. Some people like the idea
that the model's diagram (though not its equations) can be seen on the
screen, though, in practice, for even a moderately sized model, only a
small portion will be visible on the screen at anyone time.
All the packages are reasonably well equipped with built-in functions
but they are by no manner of means identically equipped . Exactly what
functions does the package support and exactly what do they do? How
cumbersome and time-consuming would it be to have to write one 's way
round a missing feature?
All the packages have to translate the model into another language .
Some packages interpret the model into , say, C, which is quite quick but
produces a model which runs relatively slowly. Other packages have an
option to compile into , perhaps, C or FORTRAN, which takes a few
seconds but produces models which run much faster than interpreted
models, which will be important if many runs are to be made or optimi-
zation searches are to be done . Is the cost of the interpretation or compi-
lation software , and a valid licence, included in the price? If not, which
compiler is recommended, and can it be guaranteed always to work
efficiently, especially regardless of the size of the model?
Some suppliers will provide a demonstration disk which will give some
insight into the product. A much better source is the manual for the pack-
age and it would be reasonable to ask the supplier to sell an inspection copy,
deducting the cost from a subsequent purchase (it would be unreasonable
to expect suppliers to give free copies of expensive publications to people
who only might buy the software) . Studying the manual before making a
commitment may save a lot of subsequent grief and the cost of buying a few
manuals would, in the long run , be trivial.
When studying a manual do not simply browse quickly through it to see
what features the product supports and what it does not , though that is the
first important test. The second test is to read at least parts of the manual
carefully to see if what is described is clearly described in a style you feel
comfortable with. Writing manuals (and books) is very difficult and it is
hard for authors to find a style which will appeal to everyone. In other
words, try to assess whether you will be able to understand the package in
quest ion well enough to use it effectively . Search for warnings about things
the software does not do; you also need to be able to use the software
correctly. The third test is to see if the manual is self-contained in the sense
that, if you understand the principles and techniques of system dynamics,
you will be able to understand the manual without any additional
knowledge. A package which required knowledge which is not in books
such as this or in its manual would be very difficult to use effectively and
correctly.
Nothing in the foregoing paragraph implies that software vendors are
Summary 373
careless or inaccurate in writing their manuals. It is simply a recognition of

the fact that writing manuals is very difficult. The manual's author knows
the software deeply and intimately and it is hard for him to avoid all the
difficulties which might be encountered by someone trying to learn how to
use the package.
Studying the manual is important but, if possible, talk to people who have
used the package you are considering and, best of all, talk to someone who
has experience in system dynamics modelling and has used more than one
package .
In general, choose carefully before buying. Consider buying more than
one package so that you can get exactly the mix of strengths you need. Once
you have developed skill in system dynamics by using one package, you will
soon learn to work in another style, switching easily to the one most appro-
priate to the purpose of the particular model you are building.
Once you have acquired the software of your dreams, do not, on any
account, rush into building big models. Work up slowly to using all the
facilities it contains, making sure you understand each one before you use
it in a serious model. For instance , suppose software X supports built-in
function Y. It would be foolish indeed simply to use Y without first having
written a small model , perhaps only a few lines long, which shows what Y
does under a variety of conditions, carefully checking by printed output,
probably at intervals as small as DT, that what Y does is what you expected
it to do. If it does not, it is extremely unlikely that the fault lies in the
software ; it lies in your lack of understanding of Y and, if you don 't under-
stand Y well enough to use it correctly and in the circumstances in which it
is intended to be used, your model is very unlikely to do the same things as
the real system and for the same reasons, and is, therefore, practically
valueless, if not downright misleading to your clients.
APPENDIX B
System dynamics problems

and solutions
INTRODUCTION
This appendix is intended to help the reader to develop skill by practice on

problems . The problems are numbered from 1, but are also arranged ac-
cording to the chapters of the book. All the models appear on the disk as
SN.COS, where N is the problem number. They are all in the traditional
DYNAMO-like format of COSMIC, and tutors may have to translate them
into the format of whatever software is available. The student should, of
course, do at least some of the problems before proceeding to the next
chapter.
In most cases, the solution to a problem will be found on the model disk,
so the solutions section of this appendix is sometimes no more than a few
words of explanation. Some of the problems have no solution.
Diagrams in this appendix are numbered according to the problem.
PROBLEMS
Chapter 1
Problem 1
Write notes to explain to another person the purpose of system dynamics
and the type of problem which it does, and does not, address.
Chapter 2
Problem 2
In a manufacturing company, the inventory of parts is depleted by con-
sumption . Consumption is determined by exogenous factors. Inventory is
replenished by the production of new parts.
The rate of ordering new parts is determined by the difference between
desired and actual parts inventory , with the rate of ordering being such as
Problems 375
to correct the discrepancy over TCI weeks. The desired inventory is the
consumption rate, averaged over a period of TAC weeks, and multiplied by
the number of weeks for which average consumption is to be covered by
inventory, TTCAS. The rate of ordering new parts is also influenced by the
average consumption rate itself.
The rate of ordering new parts feeds into a parts manufacturing backlog
and the manufacturing department set the production rate so as to elimi-
nate the backlog in TEBLOG weeks.
Using the list extension method to get started, draw an influence diagram
for this problem.
Problem 3
Oil companies import large amounts of crude oil into Europe by means of
a fleet of oil tankers. Most companies own some tankers, but the great part
of their needs is met by chartering vessels on the open market. A charter
usually lasts for between 3 and 5 years, during which time the chartered ship
fetches cargoes, mainly from the Middle East producers, as required by the
company.
The demand for refined products varies with the business cycle and is
strongly seasonal. Demand is met from refined stocks, which are replen-
ished by refinery production. Producing a ton of refined products requires
more than a ton of crude oil. For simplicity, assume that target product
stocks are constant. The rate of production in the refinery is planned to
correct any discrepancies between target and actual product stocks and to
match current average consumption.
Assuming , again, that target crude stocks are constant, the company
determines its required rate of inflow of crude oil so as to correct the crude
stock discrepancy over a suitable time period and to make up for current
average consumption of crude oil. The amount of oil which the tanker fleet
can carry at anyone time is measured in deadweight tons (DWT), and the
required shipping capacity is determined by the rate at which crude oil
should be flowing into the refineries.
Tankers are chartered so as to correct the discrepancy between the
required and actual carrying capacities over a suitable period.
Using the entity/state/transition method, produce an influence diagram.
Chapter 3
Problem 4
In the United Kingdom, a building society is a finance company which
specializes in accepting deposits from savers and providing mortgages to
house buyers. The following is a very simplified description of the mecha-
nisms operating in large building societies.
376 Appendix B: System dynamics problems and solutions
A building society is affected by three exogenous factors: the general

level of deposit interest rates in other financial sectors; the normal rate of
inflow of new deposits from small savers, and the normal demand for
mortgage lending. In practice, these three factors are not mutually inde-
pendent, but we shall assume that they are.
Funds flow into a society from several sources: the repayment of capital
from the mortgages it holds, interest payments on mortgages and the inflow
of new deposits. The last is lower when building society deposit rates have
for some time been low relative to the general level of interest rates and
vice versa. Funds flow out of a society as it lends money on new mortgages
and as it pays interest to its depositors .
The rate at which the society is willing to grant mortgages depends , in
part , on the amount of money currently held in the society. However, the
society will not be willing to lend that all at once, but will try to spread the
available money over a certain period of time; a 'cover period'.
The same 'cover period' of the normal demand for mortgages governs the
funds the society thinks it desirable to have, and the difference between that
and the actual funds drives the society's choice of the interest rate to be
offered to depositors.
The interest rate charged on mortgages is usually a multiple of that
offered to depositors. The normal demand for mortgage lending is affected
by the average mortgage interest rate and thus determines the current
demand for mortgages. The rate at which mortgages are actually made
depends on this current demand for mortgages and the rate at which the
society is prepared to lend money.
Starting with list extension, draw an influence diagram for this problem.
Problem 5
(a) Find some negative feedback loops in the solution to Problem 4 and
explain how they affect the dynamics of a building society.
(b) Repeat (a) for a positive feedback loop.
Chapter 4
Problem 6
Using your own words, explain the differences between level, rate and
auxiliary variables.
Problem 7
Write notes to explain in your own words, starting from the basic principles
of dynamic simulation, why the equations in system dynamics models can
be written in any order.
Problems 377
Chapter 5
Many of the following problems ask you to formulate equations. Always
run a simple model to test your equations and, within the limitations of
your package, thoroughly test the model by printing out all variables, and
plotting graphs of a selected set of variables. If possible, check the dimen-
sions of your equations by software; failing that , do it by hand. Ensure that
your equations do the same thing as the real system and for the same
reasons.
Problem 8
This problem is designed to give some practice in writing some very simple
equations; in some cases a few equations may be needed. Do not neglect it;
much experience shows that one of the greatest hurdles to overcome is
writing one's own equations for the first time.
Write equations for the following:
(a) Consumption Rate starts at 100 units per week and increases to 150 at
the start of week 15. It falls back to 25 units per week at week 25.
(b) Order Backlog is increased by the Rate of Receipt of New Orders and
is depleted by the Rate of Starting Production. The rate of Starting
Production is the Order Backlog divided by a time constant called the
Backlog Elimination Time.
(c) Average Consumption is the Consumption rate averaged over a period
called the Consumption Averaging Time, which is 6 weeks.
(d) The Production Start Rate depends on the difference between Desired
Stock and Actual Stock, with the discrepancy being eliminated over a
Stock Correction Time of 8 weeks.
(e) Production Completion Rate is a third-order delay of Production Start
Rate, the delay being 6 weeks. How would the equations be altered if it
was decided that a sixth-order delay of the same total duration was a
better model of this process?
Chapter 6
Problem 9
Using the Consumption Rate equation from Problem 8(a), write a complete
working model for Problem 2. Set TEBLOG, TAC, TTCAS and TCI all to
6 and use a time unit of weeks.
Problem 10
Develop a working model for the tanker chartering system in Problem 3.
Assume that product consumption has a base value of 10 X 106 tons per
month and oscillates with an amplitude of 1 x 1()6 tons per month. Target
Product Stocks are 30 X 106 • Making a ton of refined product requires 1.2
tons of crude oil so Target Crude Stocks are set accordingly. The value of
CONVF is 0.4266 and all time constants are 3 months. The duration of
tanker charters can reasonably be modelled as a third-order delay with an
average magnitude of 48 months.
Chapter 7
No particular problems are set for this chapter. You already have numerous
models to work with and you should attempt to experiment with some of
them, as suggested in the chapter.
Chapter 8
Problem 11
Write an essay explaining in your own words the principles of system
dynamics optimization, considering carefully the limitations and advan-
tages of the approach.
Chapter 9
You will be able to do some of these problems even if you do not have
access to optimization software.
Problem 12
Amend the equation for COST in the combat model optimization (p. 281)
to allow for different marginal costs of adding and subtracting resources.
Problem 13
Making whatever assumptions you wish about the marginal costs of increas-
ing and decreasing assets, optimize the model again .
Problem 14
Modify the objective function for the simple combat optimization to allow
for two constraints: money and manpower.
Problem 15
Optimize the simple combat model to achieve the minimum cost for a given
level of performance. Assume that the aim is for Blue to have a certain
number of men surviving at the end of the combat.
Problems 379
Chapter 10
Problem 16
Amend the equations for discrete production planning to allow for produc-
tion to change only at certain intervals.
Problem 17
Chapter 10 referred to the problem of making samples happen at variable
intervals , though the example given there used only two possible values of
the sampling interval. Write and test a model which takes as input a sine
wave with a base value of 100 units, an amplitude of 30 units and a period
of 25 weeks. The sampling interval has a base value of 2 weeks but varies
with random noise between 1 and 3 weeks. Such techniques are required
when one is testing the sensitivity of a system to random effects.
A technical hint is that the standard SAMPLE function takes a sample at
intervals of INTV weeks. When each sample is taken, the value of INTV at
that time is stored and a count is kept against TIME to determine when the
next sample is due. This is true whether INTV is, as usual, a constant or, as
required in this problem, a variable .
A second hint is that INTV, or whatever you choose to call the sampling
interval, must be treated as a level to act as memory for the sampling
process.
This is a nasty little problem which is intended to give disciplined practice
in thinking in terms of time as well as developing a useful technique. The
reader will gain no benefit at all by simply looking at the solution, and the
only way to approach it is to write and test equations, observing very
carefully indeed exactly what happens to the variables every DT in the run .
It your software does not make it easy to have printed output it is unlikely
that you will solve the problem correctly.
Problem 18
The bidirectional multiplier has the form:
a desfam .k= table(ttdfam,avinc.k,argl ,arg2,arg3)
d desfam = (children/family) desired family size
d avinc=($/year/family) average family income
d ttdfam=(children/family) table of value for desfam as affected by
x average family income
and ARG1, ARG2 and ARG3 are numerical arguments for the range and
step size in TTDFAM.
This formulation has been used in some system dynamics models of
economic development models to make families have fewer children as
their income increases, but it makes family size increase again if they
subsequently become poorer.
Amend the equations to ensure that, once family size has fallen due to
increased income, it will not subsequently increase, no matter how far
incomes subsequently fall.
Problem 19
Amend the air sorties model to allow for raids to take place every 2 hours
or as soon as aircraft are available and thereafter every two hours.
Problem 20
Formulate a model for the pharmaceutical company problem shown in Fig.
3.12. Assume the following numerical datal:
Average Cost of Product Research 61 $/product
Delay in Product Research 20 months
Shelf Life of a Developed Product 10 months
Delay in Product Introductions 10 months
Average Cost of Product Introduction 560 $/product
Lifetime of a Product at Market 20 months
Disposable Revenue per Product:
60 $/month at age 8 months or less
37 $/month at age 10 months
20 $/month at age 12 months or more.
Company policies are to:
• adjust product research effort over 6 months
• adjust discrepancies in products at market over 3 months
• as the average product age rises above 10 months there is a slight bias
towards allocating money to product introductions and conversely if
average product age is less than 10 months
This problem will require the use of many of the techniques covered in the
book, particularly multipliers, discussed in Chapter 5 and considered in
detail in Chapter 10. The problem is also typical of real cases in that not all
of the policies are explicitly identified.
To derive initial conditions, assume that the flows of products in the firm
are balanced at 15 products per month.
One major difficulty is the development of a formula for Average Prod-
uct Age, AP A. One approach is
APA=PY/PAM
1The data are completely fictitious and the problem is heavily adapted from a real case.
Assume all monetary values to be in tens of thousands .
Solutions 381
where PY is product years of total age, or the number of products at a

certain age, multiplied by that age and summed over all products, and PAM
is the number of products at the market. Differentiating with respect to
time (represented by a dot over the variable) yields
APA = PYIPAM - PY*PAMIPAM2
Now, PY=PAM-PWR*PLT, where PWR is Product Withdrawal Rate at
the Product Lifetime , PLT , and
pAM=(PIR-PWR)
where PIR is the Product Introduction Rate. Hence
APA=l-(PWR*PLT+(PIR-PWR)*APA)IPAM
All that remains is to translate this into system dynamics syntax.
Chapter 11
Problem 21
Choosing a problem in which you are interested, carefully identify who you
think might be the clients for a study of it. Draw an influence diagram of the
main features of the problem. Explain how you would apply system dynam-
ics to it, indicating what you would anticipate to be the most difficult aspects
of developing a quantitative model. What would you expect the qualitative
and quantitative models to tell you about the problem that you did not
already know?
Chapter U
Problem 22
Explain the main stages of the system dynamics methodology, drawing your
own diagrams to explain the points you make .
Problem 23
Summarize the limitations and strengths of system dynamics as an approach
to understanding complex managed systems.
SOLUTIONS
Solution 2
This problem is intended to give some practice in basic influence diagram
techniques. The solution in Fig. S.2 is fairly straightforward, provided one
2 COSMI C has the option of plottin g a line to mark the zero point on an axis, which will make
it obviou s if a variable has gone negative.
~ Coroumpl\on
+ Inventory - C~N5n
olP"'"
uv
Bacldog
Ellmlnalfon
Time
TEBLOG , Consumption
' \ AvP:~g
\ ~C
Produefton /
~e .......+
, /
/
, +t ~
" I Average
\ I Consumption
\ I Weeksot rate
\ I ~~ /"AVCO~
1 I In Invenloty I
11 TTCAS I \
I 1 \ I \
I I \ I
I I \ I I
, I \ I I
\ 1
/
/ I\I ~o.ir~
Inventory
I
,
/ I DtlW,
Parts
Monulacturlng
//
/ \
\
\
II" I
/
Invenloty
1 ~ \Y~'~>/
\ Rateol - - _ ..... ..... ..... "'"
Corr~t1on ... ~~ .... ---
TCI RQNP +
Fig. S.2 Influence diagram for problem 2.
carefully parsed the narrative. Check your solution against this to make
sure you correctly used solid and dotted lines, that your signs are right and
that parameters have been properly identified.
Solution 3
This is rather more difficult than the previous problem. The first trick is
to recognize that there are three entities, refined oil, crude oil and oil
tankers. The first two involve only simple levels, the third is a delayed flow.
The solution appears in Fig. 5.3.
The second key to this problem is to recognize the need for a conversion
factor between the amount of carrying capacity in the tanker fleet, meas-
ured in DWT, and the rate at which oil should, and does, flow in, measured
in Crude Tons/Month. Using round figures, the return trip between Europe
Prod uc t
+ ConSlftlP tion
_ _ _ _ -.. _.
- -- ~~ ~~~~f + ~~;t
(Pl ON/ MONTH) PSTOCK~ Aqustmen
(PTON) ,/ rima t
I " PSAT
I \ " " (MONTH)
\ , "
\ -. Pf~fi~c~n ¥ ,/
\ " + Rate Refin8d to
\ -, /,'" RPQ C I'Ude
\ ....... ...... _ .,, """ (Pl O N/ M ONTH) Conversion
, Target - . Product ,.. + 1 \ ./ Ratio
, Product stock ",. /' ,/ RCCR
" Stocks - - - - . DIscrepancy I \ ,/ ,. lC rON/PTO NJ
...... + Ave ra g e TPS + PSDIS / \ ;'
...... _ Prod uc t (p TON) (PION) / \ ,/ ,..,
JIooC~tlon / / ~+ . :
.. A PCR - - _ " Crude stcce . /
/' /' (Pl O N / M O NTH) - _ _ _ - ." C onsum ption AvafKagng
Time for
--- --- C rude
Averagi ng ,.. /' " (CTO~JS~NTH) ..... Consu mp tion
TIme ;' " rACe
TAPe " (MONTH)
(M ON TH) , /
, I
\ I
\ I
\ I
+. tI
Avofoge
Ac tual _ _ _ + C rud e Oil C rude O U
Consum ption
- Shf:>pln g;",,, -
COpaClty - - .(ClON/MONTH)
Ar~gA~f9 ;- - - - --- - - - ----- - - - - - - . . +1 C rud e Rate
A SC ... Stocks A C CR
~ (OWl) " " + C STOCK (CTON/ MONTH)
Charter
Completion ..... " (CTON)
\ 1
I
~~~
(OWl/ M ONTH)
+\' " "',
" "
\
\
I
I
+ "" "SlIpping to \ Crude I
" 0 1 Innow '\. Stoc k
" '\. conV~~JoctOl -, _ -;'DjSc~sg~ncv I
, IIC TON/M~H}!D·· - ~ .~ . _ • ..
I
' WI J Targ et I
C rude
" -, Stoc ks I
-, TCS I
,,,
(CTON) /
/
Tankef '\ "" " /
ChCJrtemg +
+ _/ _/ "
Rote ,, _ ~ " _ R~~~:" .- _
TC R
_ .. (OWl / MO NTH) _ __Sh~;g + " Re quire d - OIIR~~~W "';
SGA P ... - - - - _ Shlpplng + _- - RC SIR .. _ C rude
As:J/tsr:2.?t:.e _- - - - - C[)Wl) - - - - - - c o o c cnv .... - (CTON/ MO NTH) $lode
SGA T RSC - _ _ Aet'~~en t
(MONTH) (OWl)
CSAT
(MONTH)
Fig. S.3 Th e oil tanker charterin g probl em.

and the Middle East takes 2 months, so a ship of, say, 250000 DWT brings
in 250000 tons of crude oil every other month. Averaging over many
ships, the rate of inflow of crude oil is half the tonnage on charter. The exact
figure is 004266, so it is clearly important whether one multiplies or divides
by that number, so we have shown + and - signs on the links from that
parameter.
Solution 4
This problem is very much more difficult than problems 2 and 3. One of the
main sources of difficulty is the terminology. The 'rate' of interest is not
the same thing as a rate variable in system dynamics. To add to the confu-
sion, phrases such as 'the level of interest rates' are used and sound like
nonsense to a system dynamics ear . In fact, what is meant is the 'average
interest rate'.
The solution is shown in Fig. SA. As always, the main trick is to 'think
physics' and identify the flows in and out of the pool of funds currently held
by the society. After, that, it is a matter of mapping the influences that those
flows exert and which are exerted on them . Thus, the Rate of Inflow of New
Deposits is the Normal Rate, scaled up or down, according to whether the
society's deposit rate is, on average, higher or lower than general deposit
interest rates .
A similar idea is used for the Actual Rate of Lending. The Current
Demand for Mortgage Lending is the Normal Demand, modified by the
effects of the Mortgage Rate. The smaller of that demand and the potential
supply from the society gives the actual rate of lending.
The parameter for the duration of a standard mortgage was not men-
tioned in the narrative, but is clearly necessary.
The most interesting idea is that the society bases its Desirable Level of
Funds on the Normal Rate of Lending , which is a variable in the general
economy and which might be rather hard to determine, rather than
on the Current Demand, which could probably be measured reasonably
accurately .
Solution 5
Figure S.5A shows four negative loops. The numbering is arbitrary and it is
not implied that Loop 1 is the most significant. The reader should work out
where these loops share common paths .
Loop 1 is negative and acts to raise the deposit rate to correct a shortage
of funds. Loop 2 inevitably reinforces whatever Loop 1 does, so that if, for
example, deposit rates are raised to bring in more cash, the mortgage rate
also rises, bringing in yet more money as mortgagees pay more interest.
When society funds are low, Loop 3 ensures that less lending takes place,
while Loop 4, eventually, reduces demand for mortgages.
-- ------------ ----
--
GenefCII LfNeI
cr Decost
Interes t Rates
Gl O IR
~/Y£AR)
NormalRate
ot Irtlowof
N"'1.g~"'
(£/YEAR)
Qoteof Inflow
ot Mort~K¥M IntBl'9$t
(;'IV!AR)
+/ ' t+
/' /
// /
/
/
/'
/'
,,
- / / /
-- ---
Fig. 8.4 The building society problem .
-------------- ~~gr;o
-- / Disaepondes
/
--- --
/
, ....Building Society
/
Deposit Rate General Level
/ orceoosn
O...oble (,%~~R) ~ _ ROti.O.otG.en.",01 + _ InterestRates
/ .. levol of • - - - - - . . Interest Rates to .. .....
- - - -... Furds...... I I BuIlding SocIety GlO<R
I
I
, .-- ,/
- - OlOF
(~
......
............ +
.r>: /
/
I
I Deposit Rate
RGTBR
rr./YEARJ
I ,. ' .. ~runds I Loop I III

I
,, (Negative) \
II (
F~ +II I
I / "
/ I - Interest +
I + Outflow Averoga Rat io
,, II
NormalDemald " -" If./YEAR
IOFlOJ ) of InlE_est Rat",
AROIR
I 'a Mortgage ~~f // ell
I ~~~l
I£lVf ARJ WEAm
~~ II Oepo5dOfs \
II I " , , I Avr~na_
I , ' \ ORAl - -
\ , (YEAR) - - _. Interest Rote
I Loop 4 \ - - - ,, - EtfectonFunds
,x ;
'(Negative), .. Cu" en' Demond en"'"y -ege
"PotlendngR:1?e Mort funds
-- -- -- -... --='~kj :~: / (1)
1~8FR
I '~O( Mortgoge BSFUNO~
PMlR .- - (£.I "IIf....
, - ~m~l / ' If./YEAR/ ... + + "-
I I£,IY{AR) / /' "- +
Averag e
Mafgoge
', Rate I /', ' Rate of Inn"""
AMR / , ', OfN"\r,~1S +
, rr./YEARl
I / Lbop3 \ If./YEAR) • -, "-
I I (Negarivc) , ',,-
I ,
I ; _' / fAHation o f , "- <,
, '8otro~s ' I
\ A~ng / oSton<1Otd I -, <, NormotRote
I ~e I "- otrtlowot
#/:1) I I (YEAR) " I New Doposils
NRIND
,,
+t I <, , If./YEAR/
+. Actual Rate Rat e ot C opdal I
BulidinaSodety 'of lendnt;;J R~t Rate of IntlCl'N
Mort~~ Rete ... _ Mort Rate AROl ( IflYEARI of Mortaoao ht",,,,t
('Il./YEAR/ ---_ os~eof I + If./YEAR)
Deposit Rate • ..... RtOMI
A.+
I MRMDR
, (IJ
(~" 1 / 1
, Loop 2 _ ,,/ .--' /
(Negati ve) Mortgages _ _ .-- _ .- _ .-- " /
I MHEUD
Hold _ --
". I • (£J
I
I
I
----J --
Fig. S.5A Negative loops in the building society problem.
Solutions 387
It could well be that these four effects over-compensate for the shortage
of funds, at which point they will act so as to reduce excessive funds. One
suspects that, since the delays in the system are relatively long, the effect
may be to reinforce any cyclical effects, such as the business cycle, in the
exogenous inputs .
Figure S.5B shows a positive loop which acts to reinforce shortages or
surpluses of funds. Whenever rates rise because funds are low, more inter-
est has to be paid to depositors, thus driving funds further down. The
society hopes that this will be more than compensated for by the influx of
new funds through Loop 1, and the other negative effects in the system.
This positive loop shows how very careful one has to be in formulating
models. Does all the cash in the society belong to depositors, or is some
of it 'profit' due to the mortgage rate being higher than the deposit rate ?
It may be that the cash funds in the society are less than the amount
belonging to depositors because their funds have been lent out as
mortgages . Legally, all the assets of the society belong to its members, but
both borrowers and depositors are members, and assets may be more than
the funds on deposit. Care would be needed in building a quantitative
model!
Are there any more positive loops?
Solution 8
(a)
CONSR.KL= BASE +STEP(STEP1,STPTM1)-STEP(STEP2,STPTM2)
C BASE=100
C STEP1=50
C STPTM1=15
C STEP2 = -125
C STPTM2=25
D CONSR=(UNIT/WEEK) CONSUMPTION RATE
D BASE=(UNIT/WEEK) BASE VALUE FOR CONSUMPTION
D STEP1=(UNITS/WEEK) HEIGHT OF FIRST STEP
D STPTM1=(WEEK) TIME OF FIRST STEP
D STEP2=(UNITS/WEEK) HEIGHT OF SECOND STEP
D STPTM2=(WEEK) TIME OF SECOND STEP
In general, is better practice , if slightly more effort, to define all constants
explicitly than to write:
R CONSR.KL=100+STEP(50,15)+ .. .
The explicit form allows more flexibility for model experiments and ensures
that all factors have been defined in words, which allows their dimensions to
be more thoroughly checked .
------ ------------- -' - ~~o
/ Disaepondes
/
, /
Building Sociefy
Deo<able ~~~ate _ General lwei
" or Oeoost
+ level of .. ('I./YEAR) ~ - - - __ ~ - Gen.
Ratio. ct ef ol + _
_ - - - - . M.nds .... Interest Rate6 to A - Int erest Rates
GlOiR
/ / -' -
- OlOf < , I / Build;ng SOciefy l'Yo/VEAIl)
, (Q ................ .. funds I ~B~ote
,
... ,. .llIsctepancV I (I)
... ... I FlJjSC LoopS ,
,, / I ( Posilive ) + ...
/
, /
,, / " - .. ~ \+
No<mal Demand Period
I ot ,.. (fIV£AR)
:IOFLOWJ of Interest Rates
AverogeRatio
,, tor Mortgage Cover / AROIR
ReQUted / (I)
, Pf1ICOV I '
,· ~=l
(£/YEAR) 1YfAR) ..
\ I Deposit",,' ,
I " \ I AWH'~g ',
,,, ,,
\ \ DRAf - - _ ~ -
\ - - - - - - - _ _ ' ,_ (YEAR) - -,.. Interest Rate
It +, -, - , ENect O'l Fu>ds
,, +
~" tnnON
··,,
, LJT8nt Demand P,,=~~e -, _$:~+ ,/ I~?>FR
. . . ------_C...'0( Mort00Ql9
·,
~~ ", (~ - (~ "
--
I .. .. ,, '
·,,
I
,..
Average
(£/YEAR)
,,
,,
~ I
"- '
, AMI? , I
"" ~~ ~~
otN8W~ts
l'Yo/VEAR) I
,
..,
+,~ , I ,: (J:8.rl) -:,
, I
.,
, ,, ,,
I " , 8on'OW6t1 ' I Duatlon of ......
~ , A IJfJfOfino I a Stondard "-
.. No<mal ~e
~ ,' ::m : ~ -,
otWlowct
,,
\ ,' (YEAR) : lYE.AR) .... "-
." ,. , N"'N~
++ . . ... ~ (£!YEAR)
+~ I' .. AduOI Rate Rate of Cq::lIta
Building Sode!y +·ct~~ R~ Rate c:A rtIow
Mort~~e. ~~~e of Mort~[rrtere5t
('t./YEAlll Deposit Rate (£/YEAIl)
./~ ,.
",, T ...
... '"
MHELD - - - - - -
\"~~d£7' + II:)
- ' ,
Fig. S.5B Positive loop in the building society problem .

Solutions 389
(b)
L OBLOG.K=OBLOG.J+ DT*(RRNO.JK - RSP.JK)
NOBLOG= . . .
R RSP.KL=OBLOG.K/BET
CBET= . ..
D OBLOG=(UNIT) CURRENT BACKLOG OF ORDERS
D RRNO=(UNITIWEEK) RATE OF RECEIPT OF NEW ORDERS
D RSP=(UNITIWEEK) RATE OF STARTING PRODUCTION
D BET=(WEEK) TIME CONSTANT FOR WORKING OFF
X BACKLOG OF NEW ORDERS
A suitable initial condition would be required for OBLOG and a numerical
value for BET.
(c)
L ACON.K=ACON.J+(DT/CAT)(CONSRJK-ACON.J)
N ACON=CONSR
CCAT=6
D ACON=(UNITIWEEK) AVERAGE CONSUMPTION RATE
D CONSR=(UNITIWEEK) CURRENT CONSUMPTION RATE
D CAT=(WEEK) TIME CONSTANT FOR AVERAGING
X CONSUMPTION RATE
(d)
R PSRKL=SDIFF.K/SCT
A SDIFF.K= DESSTK.K - ASTOCK.K
C SCT=8
D PSR=(UNITIWEEK) PRODUCTION START RATE
D SDIFF=(UNIT) DIFFERENCE BETWEEN DESIRED AND
X ACTUAL STOCKS
D DESSTK=(UNIT) DESIRED STOCK
D ASTOCK=(UNIT) ACTUAL STOCK
D SCT=(WEEK) STOCK CORRECTION TIME
Note the use of the auxiliary variable, SDIFF, the value of which could be
printed or plotted as a check that the model was working correctly. Obvi-
ously, these simple equations are going to work correctly, but they may
reflect incorrect operation of some other part of the model and giving
oneself as much diagnostic help as possible is a good way to avoid a lot of
trouble.
(e)
R PCRKL=DELAY3(PSRJK,PDEL)
CPDEL=6
D PCR=(UNITIWEEK) PRODUCTION COMPLETION RATE
D PSR=(UNITIWEEK) PRODUCTION START RATE
D PDEL=(WEEK) PRODUCTION DELAY
The sixth-order delay can be represented by introducing a dummy rate,

DUMR, and writing:
R PCRKL=DELAY3(DUMRJK,PDELl2)
R DUMRKL=DELAY3(PSRJK,PDELl2)
In all the above we have referred to the entity involved as 'UNIT' . That is
acceptable for simple practice problems but, in real models, it is far better
to call the entity what it is, such as WMS for washing machines, or STTONS
for tons of steel.
From this point, you will have to run the solution models to produce the
results. This is to encourage you to gain the invaluable practice of working
with models, as opposed to looking at ouput diagrams. For example, Figure
S.9 is referred to in the solution, but you will have to run the model to
produce it.
Solution 9
The solution appears on the disk as S9.COS, though the base case behav-
iour in the upper half of Fig. S.9 should be studied first. Clearly, something
is very wrong, as the backlog of parts, PMBLOG, becomes negative at
about TIME=35, and a negative backlog is clearly ludicrous. The reason is
that non-negativity constraints were not built in, and the system's policies
are not able to handle the extreme shock of the drop from a consumption
rate of 150 parts/week to 25.
Again, this illustrates how very careful one has to be with models. It
might well have been decided that the only variables of interest were
CONSR and INV, so if PMBLOG had not been plotted the graphs of those
two variables would look plausible. It is only when one studies the printed
output that errors such as this can be found.
The reader should experiment to find a value of STEP2 which does not
cause negative PMBLOG. That will be the value which is the limit of the
robustness of the stated policies.
Turning to the model itself, none of the equations should present any
difficulty. Note that all variables are to be printed. The equation for PR is
set up, for explanatory purposes, to allow one to switch between the stated
production policy and one which includes a non-negativity constraint.
When the model is run with the latter, the graph in the lower half of Fig. S.9
is produced. If, of course, STEP2 is such that a correct response is obtained,
it does not matter whether MODSW is 0 or l.
Study of the printed values shows that INV and PMBLOG settle at a
value of 125 at the end of the run. Is that correct?
Solution 10
Most of the equations in SlO.COS are fairly straightforward, providing
one works carefully. A good example of the mistakes one can make in
Solutions 391
modelling is that, in developing this model, the author made the typing
errors:
L ASC.K=ASC.J + DT*(TCRJK -CSCR.JK)
N ASC=IASC
N IASC=ACCR*CONVF
Compare with the model to find the correct equations. In both cases the
error was found by COSMIC's dimensional analysis software . It could
have been found by hand checking, though that would have been more
laborious .
Note the inclusion of a mass balance check for tankers chartered and
develop similar check equations for refined products and crude oil. When
the model is run, the value of TCHECK is about 200, which is near enough
to zero when it is realized that cumulative charters amount to some 450 X
106 DWT over the 200 month length of the simulation.
Is 200 months too long? Probably. The period of the oscillation in con-
sumption is 51 months, so a LENGTH of about 100 or 150 should be
enough to show the main dynamics.
The behaviour in Fig. S.lO is for two cases; the base case in the main
model and an alternative of charter durations of only 12 months and the
company attempting to close its shipping gap over 12 months instead of 3.
This is obviously wrong again, as CSTOCK goes negative, reaching a mini-
mum of about -8 X 106 tons.
The reader should try to work out why this happens before reading the
next paragraph. Close study of the graphs, comparing the times at which
peaks and troughs occur, will help.
The reason for the erroneous behaviour is that large amounts of shipping
reach the end of their charters during the first few months of the run. The
company's slower reaction to placing new charters means that shipping
does not reach its peaks until about 10 months after the peaks in the base
case; the oscillations in tanker capacity have been shifted in phase relative
to those in refined stocks (and therefore relative to those in crude stock
consumption) . For this reason, the oscillations in crude stock are much
larger than in the base case and the slower chartering means that the large
fall in crude stocks in the early months is never made up. When SGAT is
left at 3 months, with the shorter charters, the behaviour becomes accept-
able.
Clearly, negative crude stocks should not happen, and it is tempting to
add a non-negativity constraint of the type:
A IPRK=PSDIS.K/PSAT+APCRK
R RPRKL= MIN(IPRK,CSTOCK.K!DT)
D IPR=(PTON/MONTH) INDICATED PRODUCTION RATE
IPR is the rate at which production should take place and RPR will be that
rate or the available stock, whichever is less.
In practice, we should adopt the strategy of regarding the negative stock

as simply a phenomenon of a catastrophic policy which should never be
adopted.
Solution 12
For the men, the required change is:
n cost=max(O,iblue- biblue)*upcost+max(O,biblue-iblue)*downcost
If the value of IBLUE chosen in a given iteration exceeds the value,
BIBLUE, from the base case, then the first MAX will operate, and vice
versa. The extension to the rest of the equation is obvious.
Solution 14
This is a relatively trivial problem.
The original objective function was:
a of.k= bpi.k/scale-lOe06*max(0,cost- budget)
Simply add another term :
-lOe06*max(0,men -menlim)
where MEN is the total of IBLUE and IBLRES, parameters which will be
varied in the optimization search, and MENLIM is the total number of men
available.
If there are two manpower limits, those available for the regular force
and those available for reserves, the penalty terms become:
-10e06*max(0,iblue-reglim) -lOe06*max(0,iblres-reslim)
This simple approach can be extended indefinitely; the scope of optimiza-
tion is restricted only by the analyst's insight!
Solution 15
The technique is to define a target to be achieved, such as Blue having
20000 men left at the end of the battle. The objective function then
becomes:
a of.k=cosH 10e06*max(O,tblue- blue.k)
c tblue=20000
The equation for COST will need to be changed so that it is now of the
form:
n cost=iblue*costreg+iblres*costres
and so forth , as all we are concerned with is the minimum cost of a given
Solutions 393
performance, not the maximum performance for a given budget. The

method of the previous solution can be used to avoid overspending.
The objective function, OF , is now to be minimized, as the problem is to
achieve a given performance for minimum cost. If the optimal solution is to
reduce current levels of spending, COST will be negative, which will be
ideal. If BLUE.K, the value of Blue's force at the end of the simulation, is
less than TBLUE, a huge penalty will be added to OF and the minimization
will not be achieved . On the other hand, if BLUE.K is greater than
TBLUE, no penalty will be added.
That could, of course, leave Blue with more than 20000 men at the end of
the battle, which, in human terms would be fine. In the brutal world of
optimization, one might aim to be as close as possible to 20000, in which
case one might use:
a of.k=cost + 1Oe06*(tblue - blue.k )**2
The method can be varied to make BLUE follow some desired profile
throughout the battle and penalizing variations below, or from, that
profile.
There is, of course, no magic about the use of lOe06 as a penalty weight.
If, for example, COST is of the order of lOe09, then the penalty weight
needs to be lOe12.
Although these two solutions are in terms of the simple combat model,
the ideas apply to any optimization problem.
Solution 16
This is very straightforward. Instead of using:
r pr.kl = opr.k +pchan*int((dpr.k - opr.k +0.5*pchan)/pchan)
we write:
a dum.k= opr.k + pchan *int((dpr.k-opr.k+ 0.5*pchan)/pchan)
so that DUM can be printed out as a check that the equations are working
properly and then use:
r pr.kl=sample(dum.k,prodper,dum.k)
d prodper=(week) interval at which production planning decisions
x are made
Solution 17
* Figure s.l7 variable sampling times
note
note file named s17.cos
note
note sine wave to be sampled

note
a input.k = base + amp*sin(6.283*time.k/perd)
c base=lOO
c amp=30
c perd=25
A standard sine wave for INPUT, the variable to be sampled.
note
note sampled value of input
note
a output.k =sample(input.k,intv.k,input.k)
The resultant sample is OUTPUT.
note
note sampling interval sector
note ============
note
note noise in sampling interval
note
a samvar.k=int((basint*(l +noise(seed))+O.5*dt)/dt) *dt
c basint=2
c seed=19
SAMVAR is expressed as an exact number of DTs. There is no need to take
random numbers at set intervals, as was done in Chapter 10, as all we need
is to store the appropriate value of SAMVAR when INTV is changed.
note
note saving of random value until a sample is due
note
I save.k= save.j+ (dt/dt)*samvar.j*clip(1,O,save.j- time.j,dt)
x *clip(1,O,dt,save.j- time.j)
n save=basint
This is the key sector of the model. The sampled value is saved to monitor
the passage of time. Since, however, the sampling interval will only be
changed when a sample is taken , we require to store the value of SAMVAR
1 DT before the sample is due to be taken. This will then be held as the
sampling interval to determine when the next sample is due.
The two CLIPs ensure that the change only takes place when the differ-
ence between SAVE and TIME is exactly 1 DT. Run the model again,
removing one or other of the CLIPs, and explain what happens.
note
note saving the sampling interval at the appropriate time
note
Solutions 395
I intv.k =intv.j + (dt/dt)(samvar.j -intv.j)*clip(l,O,save.j-time.j,dt)
x *clip(l,O,dt,save.j- time.j)
n intv= basint
The value of INTV is updated whenever the comparison between SAVE
and TIME makes TIME 1 DT before SAVE.
note
note model control and output
note
c dt=O.25
c pltper=O.25 for precision graphs
c prtper=0.25 printing every dt to check model!!!!
c length=40
print l)input
print 2)output
print 3)samvar
print 4)save
print 5)intv
plot input = a,output = b(O,200)/intv=c(O,5)
This sector is absolutely critical. Note that each variable is printed at
intervals of DT to ensure that the right things are happening at the right
times. Only a few variables are plotted, though simply plotting INPUT and
OUTPUT might suggest that the model is correct. Figure S.17 shows that,
as required, INTV changes exactly when the samples are taken, except
where the randomness has produced the same value for INTV twice in
succession.
note
note model tests
note
run A trial of equations
c seed=3
run B a different series of random numbers
Testing the model with at least two sequences of pseudo-random numbers
increases confidence that it is working correctly .
This procedure of testing the tricky parts of models by writing a mini-
model and closely studying what is happening is an essential method of
developing anything other than a purely trivial case.
note
note
d amp=(unit) amplitude of sine wave for input
d base=(unit) base for sine wave for input
d basint=(week) base value for noise
d dt=(week) time step in simulation

d input=(unit) variable to be sampled
d intv=(week) sampling interval
d length=(week) simulated period
d output=(unit) sampled value of variable to be sampled
d perd=(week) period of sine wave for input
d pltper=(week) plotting period
d prtper=(week) printing period
d samvar=(week) values for random sampling interval
d save = (week) cumulatively saved random sampling intervals
d seed=(1) seed for random number generation
d time=(week) time within simulation
Solution 18
It is not possible to write, as one might in FORTRAN, C or Pascal,
a desfam.k=min(desfam.k,tabhl (. . .»
as that requires DESFAM to depend on itself, which prevents the package
sorting the equations into the correct computational sequence.
A possible solution is to store the previous value of DESFAM:
I odesfam.k =odesfam.j+ (dt/dt) (desfam.j-odesfam.j)
d odesfam=(children/family) value of desfam one dt previously
and then to use:
a desfam.k =min(odesfam.k,table( . . .»
Solution 19
* Figure s.19 revised aircraft sortie model
note
note file named s19.cos
note
Most of the model is unchanged from FIG10-11.COS, so only the new
equations need to be explained .
note
note launch rate
note
note the policy switch, POLSW, allows us to change from
note one decision rule to another
note
c polsw=1
To demonstrate the difference between the old policies and the new, a
policy switch is introduced.
Solutions 397
r launch.kl = polsw*clip(sgroup/dt,O,acavail.k,sgroup)
x *pulse(l,wstart,abintvl.k) + (l-polsw)
x *pulse(clip(sgroup/dt,O,acavail.k,sgroup),wstart ,stsint)
The first two lines are the new policy of launching every two hours or as
soon as the required number of aircraft become available . The CLIP tests
for aircraft availability and the PULSE sends a series of Is to allow the
CLIP to work (or not). The frequency of the Is is shown below and is either
the standard interval, STSINT, of 2 hours, or DT, which has the effect of a
continued monitoring of aircraft availability when there were too few to
launch at the last two hour point.
The third and fourth lines are the old policy from Chapter 10.
a abintvl.k= clip(stsint,dt,acavail.k,sgroup )
c stsint=2
The launch interval, ABINTVL, is either 2 hours when aircraft are avail-
able or a continuous check every DT if they are not.
c sgroup=40
The groups have been set to 40 in order to get some clear comparisons
between the two policies. The reader should, of course, experiment with
other group sizes.
note
note record time at which launches took place
note
I ltime.k= ltime.j + (dt/dt)( time.j -ltime.j)*clip(l ,O,launch.jk,sgroup)
n Itime=O
An equation to record the times at which launches took place.
note
note model control and output
note
c dt=O.015625
c length=40
c pltper=O.125 for precision plots
With such a small value of DT, PLTPER must be chosen by experiment to
get reasonable graphs . Run the model with PLTPER=l and observe the
difference in apparent behaviour. This is yet another demonstration of the
inability to debug a model properly from the plotted graphs.
print 8)abintvl,ltime
The new variables are added to the print commands for debugging.
plot acavail= a,oparea = b(O,100)/ltime=c(O,50)
The launch time variable , LTIME, is added to the plots.
run A Launch every 2 hours or when 40 available

c polsw=O
run B launch every 2 hours if 40 are available
Experiments to compare the two policies.
note
note
d abintvl=(hour) minimum interval between launch of groups of
x aircraft on sorties
d launch=(aircraft/hour) rate of launch on missions
d ltime=(hour) time at which launches occurred
d polsw=(I) switch to choose alternative policies
d stsint=(hour) standard interval for launching sorties
Definitions are added for new or revised variables .
The behaviour with the two policies is shown in Fig. S.19. The behaviour
is quite significantly different, the old policy giving 10 launches over the 40
hours while the new is nearly ready for a twelfth. With the old policy, there
are some quite clear cases where rather more than 30 aircraft are available,
but nothing happens until the next two hour point is reached.
The value of developing models to include such fine detail will depend on
the case. There is no virtue in complicating a model simply to show one's
modelling virtuosity. However, policy nearly always makes a significant
difference to system behaviour, so correct modelling of policy is usually
vital.
It is an open question whether some of the packages described in Appen-
dix A would be capable of this degree of sophistication.
Solution 20
* Figure s.20 the pharmaceutical company model
note
note file s20.cos
note
note research and product availability
note =================
note
a rrsr.k = ((dpafm.k -pafm.k)/prat+ apir.k) *acpr
The aim is to bring Products Available for Marketing up to a target level
and to replace products being introduced, the effect being expressed as a
desired rate of spending.
c acpr=61
c prat=6
a aspr .k=max(O,drflow.k-aspi.k)
Solutions 399
Actual spending may be less than the desired level. The model makes no
provision for the spending of accumulated balances, an area which the
reader may wish to explore.
note
note product research starts and completions
note
r prsr.kl=aspr.k/acpr
r prcr.kl=delay3(prsr.jk,dres)
note
note products available for marketing
note
1 pafm.k=pafm.j+dt*(prcr.jk- rspi.jk- par.jk)
n pafm=base *dres
c base=15
The pool of Products Available for Marketing is initialized to the base flow
of 15 products per month.
c dres=20
a dpafm.k=apir.k*dres
The target for Products Available for Marketing.
r par.kl = pafm.k/sl
c sl=lO
Products are lost due to the shelf life effect.
note
note product introductions
note ===========
note
a rrspi.k= ((apwr.k*dlife- pam.k )/mpat + apwr.k)*acpi
c mpat=3
A similar concept to that use for Products Available for Marketing.
a aspLk=min((pafm.k/dt)*acpi,min(l,rsr.k*pisf.k)*drflow.k)
The Actual Spending may be increased by a factor affected by Average
Product Age, PISF (see below).
r rspi.kl=aspi.k/acpi
n rspi=base
r rcpi.kl=delay3(rspi.jk,dintro)
c dintro=lO
c acpi=560
1 apir.k =apir.j + (dt/tapi)(rcpi.jk - apir.j)
n apir=rcpi
c tapi=2
Simple equations of flow and averaging.

note
note products at market
note
I pam.k=pam.j+dt*(rcpi.jk-pwr.jk)
n pam=base*dlife
note
note product withdrawals
note
r pwr.kl= delay3(rcpi.jk.dlife)
c dlife=20
1apwr.k=apwr.j+(dt/tapwr)(pwr.jk-apwr.j)
n apwr=pwr
c tapwr=2
Again, a simple flow of Products at market.
note
note average product age
note = = = = = = = = = =
note
1apa.k=apa.j +dt*(l-(pwr.jk*dlife+ (rcpi.jk-pwr.jk)*apa.j)/pam.j)
n apa=dlife/2
The mathematical derivation gives the rate of change of APA and therefore
appears in a level equation as shown.
note
note revenue flows
note ===== = =
note
a drfpp .k=tabhl(trl,apa.k,8,12,2)
t trl=60/37/20
a drflow.k=drfpp.k*pam.k
I cumearn .k=cumearn.j +dt*drflow.j
n cumearn=O
note
note product development policy
note ==============
note
s tns.k=rrsr.k+rrspLk
A supplementary variable, solely for output.
a rsr.k = rrspi.k/(rrsr.k + rrspi.k)
The ratio of the two spending requirements, which was used to determine
the Actual Rate of Product Introductions.
Solutions 401
a pisf.k=tabhl(tpisf,apa.k,9,1l,l)
t tpisf=O.9/1.0/1.1
The actual ratio is scaled up or down depending on Average Product Age.
a tots.k=aspr.k+aspLk
a factor .k=tots.k/drflow.k
Some additional output variables.
note
note output and control
note ==========
note
c dt=O.25
c length=lOO
c pltper=l
a prtper.k=l + step(9,l 0.5)
print 1)prsr,prcr,factor
print 2)pafm,par,dpafm
print 3)rspi,rcpi
print 4)pam,pwr
print 5)apa,drfpp,cumearn
print 6)drflow,tns,pisf
print 7)rrsr,aspr,rsr
print 8)aspi,aspr,rsr
plot pam=a,pafm=b(O,800)/apa=c(8,12)
As always, all the variables have been printed and the output closely exam-
ined.
note
note
run A Trial of model
note t tpisf=O.O/l.O/2.0
note run very steep tpisf
note t tpisf=O.9/1.0/1.1
c prat=12
run B Policy of smoother reactions
Two experiments are carried out, though many more are suggested.
note
note
d acpi=($/product) average cost of product introductions
d acpr=($/product) average cost of product research
d apa=(month) average product age

d apir=(product/month) average product introduction rate
d apwr=(product/month) average product withdrawal rate
d aspi=($/month) actual spending on product introductions
d aspr=($/month) actual spending on research and development
d base = (product/month) base value for initial conditions
d cumearn=($) cumulative earnings
d dintro=(month) delay in product introductions
d dlife=(month) product lifetime
d dres=(month) delay in product research
d drflow=($/month) disposable revenue flow
d drfpp=«$/month)/product) disposable revenue flow per product
d dpafm=(product) desired products available for market
d dt=(month) solution interval
d factor=(l) ratio of total desired spending to available revenue flow
d length=(month) simulated period
d mpat=(month) marketable products adjustment time
d pafm=(product) products available for marketing
d pam = (product) products at the market
d par=(product/month) product abandonment rate
d pisf= (1) product introduction preference factor. A multiplier to
x show the preference attached to product introduction
x when allocating spending.
d pltper=(month) plotting interval
d prat=(month) product research adjustment time
d prcr=(product/month) product research completion rate
d prsr=(product/month) product research start rate
d prtper=(month) printing interval
d pwr=(product/month) product withdrawal rate
d rcpi=(product/month) rate of completing product introductions
d rrspi=($/month) normal rate of spending on product introductions
d rrsr=($/month) normal rate of spending on product research
d rspi=(product/month) rate of starting product introductions
d rsr=(l) ratio of required spending on product introductions to total
x required spending on research starts and product
x introductions
d sl=(month) shelf life of a developed product
d tapi=(month) time for averaging product introductions
d tapwr=(month) averaging time for product withdrawal rate
d time=(month) simulated time
d tns=($/month) total required spending
d tots=($/month) total actual spending
d tpisf=(l) table for pisf
d trl=«$/month)/product) table for revenue level
Solutions 403
The output is shown in Fig. S.20 for two simulations and in com-
parative form for each of the three variables. The main point is that ,
as ever in system dynamics, policy has an enormous effect on system
behaviour.
APPENDIX C
COSMIC supporting
documentation and software
INTRODUCTION
This appendix provides information on documentation , software and sup-

porting services available from the COSMIC Holding Company.
TRAINING
Training in system dynamics is available from the author or his associates.

It can be a one-week course, as outlined in Chapter 12, or shorter versions
for senior managers or analysts who require an appreciation of system
dynamics, rather than a full training course. All courses are designed spe-
cifically for the customer.
CONSULTANCY
Consultancy in system dynamics is available from the author, supported by

his associates where necessary. Fees are charged on a daily basis, plus any
significant expenses.
DOCUMENTATION
Several publications are available which support the use of system

dynamics.
• An Introductory Guide describes the features of COSMIC and COSMOS
in more detail.
• A ISO-page User Manual explains all procedures for using COSMIC and
COSMOS.
• Equations for Systems provides more than 30 case studies of the use of
COSMIC and COSMOS for advanced modelling.
COSMIC and COSMOS software prices 405
• System Dynamics Problems has 45 problems on all aspects of system

dynamics, with full solutions.
COSMIC AND COSMOS SOFTWARE PRICES
All prices for COSMIC and COSMOS include a licence for Prospero FOR-
TRANI, and one copy of the User Manual for each licence. COSMIC and
COSMOS are supplied on 3.5"disks as one integrated package . The licence
fee covers both packages and the royalty we pay to Prospera Software.
Payment may be made in any convertible currency, plus 5% for orders from
outside the UK. Order by official purchase order, or cheque, adding postage
and packing as specified below.
UK orders for software are subject to value added tax (VAT) ; the current
prices (1995) are shown below. The amount should be added to cheque
purchases and a VAT receipt will be sent with the software. Orders from
outside the United Kingdom are not subject to VAT. Multiple copy prices
apply only to orders placed within 90 days of the first order.
Standard licence
First copy £800 (£931.25 including VA 'f2)
Next four copies £500 each (£578.75)
Subsequent copies £350 each (£402.50)
Academic institution licence

First copy £500 (£578.75)
Next four copies £300 each (£343.75)
Subsequent copies £250 each (£285.00)
Software with an academic licence may only be used for teaching and
scholarly research and not for commercial consultancy or grant-supported
research.
Individual student licences

A charge of £250 (£285.00 including VAT) is made . Orders must be accom-
panied by a statement from the Head of Department that the software will
be used only by a named student for study and research, will not be given or
1 Supplied by Prospero Software Ltd, 190Castelnau, London SW13 9DH , UK; telephone 0181
741 8531. The licence covers only those part s of Prospero FORTRAN that are necessary for
COSMIC and COSMOS. A full FORTRAN package is not supplied and there is no FOR-
TRAN manual.
2 Prices allow for the fact that VAT is not charged on documentation.
406 Appendix C: COSMIC supporting documentation and software
sold to any other person and will not be used in the student's subsequent
employment.
DOCUMENTAnON PRICES
The Introductory Guide is free . The User Manual is £50, which will be
deducted from any software purchase made within three months. Equations
for Systems and System Dynamics Problems cost £50 each. VAT is not
charged on documentation.
POSTAGE AND PACKING
Postage and Packing for each licence or book are £5 within Europe, £15 in
the rest of the world.
GENERAL
Specification and prices may change without notice. We reserve the right to
decline orders for software and documentation from competitors. For infor-
mation or orders please contact:
The COSMIC Holding Company

8 Cleycourt Road
Shrivenham
Swindon
Wilts
SN68BN
UK
Telephone +44 (0)1793 782817

APPENDIX D
Using the disk
INTRODUCTION
The disk at the back of the book contains:

source code for 59 models in COSMIC format ;
99 influence diagrams in CorelDRAW Version 3 (CorelDRAW is not
supplied with the book);
the text for Chapter 10.
The models are all those mentioned in the book and will be invaluable to
you in developing your modelling skills through practice. For instructors
and advanced students they provide a host of opportunities for model
development and revision.
All the models should run with only minor amendments in DYNAMO
and DYSMAP2. Users of the other packages referred to in Appendix A will
have to rewrite the models but this will be marvellous practice in modelling,
requ iring close study of the supplied models and thereby developing excel-
lent understanding of the principles of system dynamics, which are the
subject of this book .
The influence diagrams will be helpful if you have access to
CorelDRAW, a very widely used graphics package , or to any other graphics
package which will accept its formats, They will allow you to develop skill
in drawing and revising diagrams which is the basis of system dynamics
thinking.
The text of Chapter 10 appears as CHAPlO.DOC for Microsoft Word
5.5, which may also be readable by other word processing packages, and as
CHAPlO .TXT, which is in ASCII format and which can be printed by the
normal DOS print command. The influence diagrams for Chapter 10 are in
the book and the models are on the disk.
READING THE FILES
The model and influence diagram files on the disk are in compressed for-
mat. To load them, insert the disk in drive A:and type
A:LOAD
408 Appendix D: Using the disk
and press enter. This will create a directory called DIAGS as a subdirectory
of C:. The compressed file of diagrams and the decompression software will
be copied into that directory and the files will then be decompressed after
which the compressed file and the decompression software wil be deleted
from the subdirectory. Similarly, the models will be decompressed into a
subdirectory called MODELS. Finally, the text of Chapter 10 will be copied
into a subdirectory called CHAP10 and will exist as CHAPlO.DOC and
CHAP10.TXT. If you do not wish to use these directory names, you can
amend LOAD. BAT on the disk as you wish. If you are not fully conversant
with DOS get someone to help you.
Users of COSMIC will have to split the models up into groups of not
more than 14 and copy them into directories under the COSMIC main
directory.
Index
Aggregated influence diagrams 40 Conserved quantities 87

Al-Alusi, A.-S. 342 Constant 92
Arrays 164 Contents of a delay 64
Auxiliary variables 84 Control engineering 211
correct relationship to levels 161 Control strategy 257
dimensions of 88 Control systems 2
Cosine function 114
Base vector optimization 364 COSMIC 14,93,361,362,363,367,
Behavioural effects in managed 369,370,374
systems 51 COSMIC User Manual 231,404
Bidwell, S.W. 340 COSMOS 14,237, 363, 371
Bright ideas 10 Counter-insurgency warfare 340
Building societies 375 Counter-intuitive behaviour 108, 169,
Business process re-engineering 197 226,347
Coyle, J.M. 330, 342, 359
Clanguage 372,396 Coyle, RO. 9,307,341
Cascading of delays 101
Case studies in modelling 140 D statements 95
Causal loop diagrams 10,18,365,366 Damping 215
Cavana, RY. 309 Debugging 369
Central heating system 18 by printed output 148, 165, 175,
Checkland, P. 350 312,373
Choosing a suitable delay function Decision theory 1
124 Definitions of variables
Clark R 342 D statements 95
Clark T. 342 importance of 153
CLIP function 111, 119, 230, 394, DELAY3 148,389,399,400
397 Delayed flow module 59
Closed systems 57, 180,231 Delays in models 98, 207
Club of Rome 307 as source of instability 208
Combat model 80, 112, 169, 222 correct balancing of 162
optimization of 277 Dimensional analysis 23, 24, 88, 116,
Combat power 68 312,356,370
Command and control (C2) systems avoidance of modelling errors 194
340 inability to detect some model
Common modules 50,58,141 errors 161
Common modules method 36 means of detecting errors 159
Comparative plots 106 software for 95,215
Computed constant (N equation) Dimensional consistency 76, 354
204 in objective functions 244
Computer Simulation in Management Dimensional equation 65, 87
Science 359 Dimensional identity 87
410 Index
Dimensional validity 65, 153 Growth-producing feedback loop 8

checking of 158 Gulf War 342
Dimensions 23, 51, 65, 76, 85, 377
Discrepancy 22, 202 Hackett, Sir John 340
Division by 0 175 Heuristic method 240
Domestic Manufactur ing Company overcoming limitations of 241
(DMe) 25, 27,140 , 196, 197
optimization 250 Identification of the potential client
Driven systems 57,180 349
DT 84, 85, 112, 373 Industrial Dynamics 306, 358
correct use of 354 Infectious disease modelling 182
incorrect , effects of 188 Influence diagrams 8, 10, 18
on the right hand side of a rate or characteristics 47
auxiliary equation 109, 116, 124 cone of 18,43, 54, 60, 76, 143, 329,
permissible values 188 352
selecting a value 109, 148, 184 conventions for 20
significance 109, 188 equivalence to equations 146
Dynamic behaviour 5 guidelines for drawing 40
Dynamica 309 layout of 28
DYNAMO 111, 135, 148, 362, 364, limitations of 45
367,369,370,374 methods for building 31
DYSMAP2 367,369 ,370,371 organization of complexity 358
right method for 39
Endogenous factors 23,116,306 roles of 81
Entit y/state/transition method 33, 375 transparency of 15
Equations for Systems 404 Information and action parts of
Eulerian integration 188 influence diagrams 20
Exogenous factors 23,41,114,258, Information feedback 2
307,376 structure 3
Experimentation Information system modelling 358
vs, design 196 Information/action/consequences
External driving force 20 paradigm 4,8, 19,20,26,28,84,
Extreme conditions test 354 147,222
Initial condition 90
False loops 71 Initial values
Feedback loop structure 217 setting up by constants 123, 173
Feedback loops 5, 21 setting up by variables 146
polarity 21 Injecting quantities into a model 116
Feedback structure 55, 60, 65, 71 Input switches 145
Feedforward 215 INT function 394
First order delay 101, .164 Interest rates 384
Flat-bed plotter output 148 Introduction to System Dynamics
FLIP function 231 Modeling with DYNAMO 358
Flow modules 81
Forrester, J.W. 3,9,306,348,355, Kill probability 68
358,359,362
FORTRAN 361,363 ,372,396 LENGTH 91,113
Prospero 405 rule of thumb for choosing 310
selecting a value 148,391
Gain 207 Level variables 84
as source of instability 208 as accountants 89
Gardiner, P.A. 341 as memories in systems 141
Goal-seeking feedback loop 7,31,55 as memory of information 141
Graphic user interfaces 368 dimensions of 88
Index 411
initial values for 146 Nesting of functions 122
used to generate memory 230 NOISE function 117
Level/rate diagrams 45, 142 Non-negativity constraint 53, 68, 116,
Limiting functions 121 123,163,174,180,390
Lim its to Growth, The 307 Non-linearities 130
Linear programming 1, 13,309,350 additional methods of representing
List extension 58 135
List extension method 31,50,375 and optimization 239
Logic in models 119 dimensions for 135
NORMD function (normal
Management consultancy firm 49, 155 distribution) 117
Management System Dynamics 360 NOTE statements 145
Managing the Energy Transition 309
Mass balance equations 96, 153, 354, Objective function 13, 237, 392
370,391 candidate or competing 352
as auditors 98 development of 254
MAX function 121,392,398 dimensional consistency in 244,
Mayan civilization 73,112 254
Meadows, D.H. 307 factors in choosing 292
Meadows, D.L. 307 formulation of 243
Military modelling 223 more complex 287
MIN function 121,391 ,396,399 scaling of 243, 256
Minimization by method of steepest significance of 244, 293
descent 244 Oil companies 375
Mode of dynamic behaviour 96 Open systems 57
Model Optimization 13, 16, 236, 269, 366,
assessing the quality 353 371
documentation 93, 136,312 as guided search of the parameters
identifying limitations 307 236
mapping 125, 136 assessing the value of 271
purpose 180 by hill climbing 239
testing 177,186 by repeated simulation 249
tests of confidence 153 constrained 249, 276, 342
wrong while appearing right 178 costly parameters in
Model list 31,50 constrained 282
Modelling experiments 244
Argentina's electricity capacity 315 experiments as an insight 273
as a political act 349 explanation of results by verbal
fleet of aircraft carriers 341 reasoning 283
fleet of submarines 341 for designing policies and structures
New Zealand forestry industry 309 236
Modelling of competition 223 for fitting models to data 295
Multipliers 380 free parameters in constrained 282
Multipliers in system dynamics models importance of experimentation
133 before 249
in practice 249
N equation 90 maximization and minimization
Naill, RF. 309 279
NCLIP function 290 methods 364
Negative feedback 7,31,202 non-linear 294
essential idea of 7 not a substitute for thought 272
laws of 202 objective function for constrained
Negative sign in influence diagrams 280
20 priorities in constrained 286
412 Index
Optimization (Continued) Positive sign in influence diagrams
problems 378 20
software 13 Powersim 362, 365, 368, 371
subtlety of results 260 Practical modelling
unconstrained 249, 342 political aspects 165
Predator/prey model 62,71 ,78,112,
Parameter plane 237,262 180
Parameters 20 Pressure points 67,79,170,222,231
analysis of 207 and optimization 277
component of policy 196 and policies 222
costly 235 experimentation with 223
free 235 remarkable effect of changing 236
remarkable effect of changing 236 Principle of physical flow 347
structural 202 Principle of robustness 6
Parsing 26, 34, 50, 58, 80 Production management problem 24
Pascal 396 Proportional control 26, 202
Pet theories 10,169,198,211 PRTPER 95,148
Pharmaceutical company problem 57, Pugh, A.L. 358
74,380 ,398 PULSE Function 115, 125, 230, 397
Physical flows in influence diagrams Purpose of a model 48
20
Pidd, M. 359 Qualitative analysis 10, 13, 54, 60
Piece-wise linear approximation for Qualitative system dynamics 323
nonlinearities 133 air traffic control system 323
Pipeline delay 101 drugs 'problem' 326
Planning horizon optimization 364 police force measures of
PLTPER 95,148 performance 331
Policies 2, 147 universal industries 330
Policy 23 Quantitative model 84
as a pressure point 228
effects on behaviour of 108,163 RAMP function 113
in system dynamics models 195 Random effects in models 117
meaning 222 fundamental errors 117
mixed 211 Rate variables 84
switches 202, 396 dimensions 88
two components 196 Rate-dependent rates 194, 354
Policy analysis 54 Ratios
two phases 197 correct representation 159
Policy design 13, 49, 102 Reference mode 49
Policy experimentation 197 Rego, J.e. 315
Policy experiments 106 Response surface 239, 262
and noticeable changes of behaviour Richardson G. 359
195 Richardson H. 342
with system dynamics models Roberts , N. 358
195 Robust behaviour 26
Policy optimization 197 Robust policies 13
Policy option diagram 256
policy level 258 SAMPLE function 116,117,379,394
strategy level 258 Significant delay 27
Policy options 53 Simple combat model 67
Policy parameter 91 Simplification
Policy switch 106 by optimization 364
as structural parameter 211 Simulation 84
Positive feedback 7 by repeated optimization 247,249
Index 413
example 91 process 13
fundamental ideas 84 purpose 315
Simulation experiments 152 qualitative 323
Simulation model role 343
ease of writing from influence software packages 14,92,111,361
diagram 12 structured approach 10
Simulation process 90 teaching 355
Sine function 114,394 water-flow analogy 84, 141
Sine wave input 145 System Dynamics Problems 405
Smoothed level 89, 141 System Dynamics Review 309, 359
Smoothing equation 103,145 System Dynamics Society 359
dimensions of 104 Proceedings of annual conferences
Smoothing of information 21,102 309
Soft systems methodology 350 System Enquiry 359
STELLNithink 362,365,367,369,
371 TABHL function 135, 158, 400
STEP function 93,113,387,401 TABLE function 130, 159, 379
Step input 145 Tanker chartering 377
Step length 84 Tests of confidence in a model 96
Step response 153 The Third World War 340
Structural analysis 210 'Think physics' 19,26,35,58, 141,
Structural parameter 211 142,347,384
binary 211 Third-order delays 101
continuous 211 Time 91
Structure TIME 112
as component of policy 196 as clock for simulation 91
remarkable effect of changing 236 dimensions 112
Study periods 15, 329, 352 equation 112
System initial 112
definition of 4 Time constant 26
immense range of design possibilities Time delay 20
231 Time label 85
System description 10, 13, 63, 81 Time shift and relabelling in simulation
System dynamics 89
applications of 306 Time-step simulation 84
appropriate circumstances 348 Two-analyst problem 33
bibliography 306, 343 Type assignment 91, 141
characteristics 15 Type label 85
criteria for suitability 309
defence analysis 340 Undriven systems 57
definition 5, 9, 359
design of management information Validation 12,153
systems 198 definition 155
difficulty of starting with a blank Variable naming 141
piece of paper 350 Variables
essence 346 rules for time postscripts 192
essential aim 347 Vectors 164
fundamental equations 85 Vensim 362, 366
history 306
identification of market 349 Wolstenholme, E.F. 9,340,342,343
in practice 350 World Dynamics 307
management science 1,16 World War 3 340
paradigm 3 WORLD2 model 307
problems and solutions 374 WORLD3 model 307
Errata
Users of COSMIC should replace the exponent **2 in the objective functions
on pages 243, 254, 255 (three occurrences), 280 and 393 with INT(2) and
make the corresponding changes in the models on the disk (FIG8-5M.COS,
FIG9-6.COS and FIG9-12.COS - there is no model on the disk for Problem
15). Users of other packages should consult their manual. Depending on the
package, an execution error may occur with X**N if X is negative and N is
treated as a real number. COSMIC avoids the problem because its INT
function converts real numbers to integers. An alternative approach is to use
the absolute value of X: ABS(X)**N. If the package does not support
exponents, just use ABS(X). If in doubt, try your package and see what
happens, or contact its supplier.
System Dynamics Modelling . R.G . Coyle.

Published in 1996 by Springer-Science+Business Media,B.V.
ISBN 978-0-412-61710-2

R.G. Coyle - System Dynamics Modelling

Uploaded by

Copyright:

Available Formats

R.G. Coyle - System Dynamics Modelling

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

R.G. Coyle - System Dynamics Modelling

Uploaded by

Copyright:

Available Formats

System Dynamics Modelling

JOIN US ON THE INTERNET VIA WWW, GOPHER, FTP OR EMAIL:

Springer-Science+Business Media, B.Y.

~ Printed on perm anent acid-free text pap er, manufactured in accord -

1 Introduction to system dynamics 1

3 Influence diagram case studies 48

The consulting firm's problem 49

5 System dynamics modelling techniques 111

6 Case studies in modelling 140

The simple combat model 169

7 Policy experiments with system dynamics models 195

8 Optimization in system dynamics 236

9 Optimization in practice 249

10 Advanced modelling 297

11 The applications of system dynamics 306

12 Summary and conclusion 346

Appendix A: System dynamics software packages 361

Appendix B: System dynamics problems and solutions 374

Appendix C: COSMIC supporting documentation and software 404

Appendix D: Using the disk 407

SYSTEM DYNAMICS AND MANAGEMENT SCIENCE

A good example of a dynamic system is the ordinary domestic central

THE PARADIGM OF SYSTEM DYNAMICS

The paradigm, or basic viewpoint and associated methodologies, of system

human being cannot be explained from the chemical composition of the

Fig. 1.1 The information/action/consequences paradigm of system dynamics.

POSITIVE AND NEGATIVE FEEDBACK

under the influence of the system's policies, so as to attempt to eliminate the

systems and may be valuable as an engine of growth. In an engineering

DEFINITIONS OF SYSTEM DYNAMICS

System dynamics deals with the time-dependent behaviour of managed

A STRUCTURED APPROACH TO SYSTEM DYNAMICS

Having discussed some of the underlying ideas of what system dynamics is

Sta ge 1 Problem Recognition

Stage 2 Problem Understanding and System Description

Sta ge 3 Qualita tive Analysis

Stage 4 Sim ulation Modelling

Stage 5 Policy Testingand Design .... ---1

(objec tive function)

Fig. 1.3 The structure of the system dynamics approach,

numb er of people available could be a symptom of that. Similarly, a system

THE PROCESS OF SYSTEM DYNAMICS

Fig. 1.4 The process of system dynamics.

We have seen several times that system dynamics concentrates on the

System viewpoint Whole company

Feedback analysis Consequences of actions

Concern with Mure

Transparency of Understanding. input

Fast s1mulal1on Study periods

Fig. 1.5 System dynamics and top management.

This chapter has been an introductory survey of what system dynamics is

In Chapter 1, we stated that influence diagrams are an essential tool in

AN INFLUENCE DIAGRAM FOR A CENTRAL

Fig. 2.1 Influence diagram for central heatingsystem.

Unfortunately, the higher the room temperature relative to the outside

THE CONVENTIONS FOR INFLUENCE DIAGRAMS

It is now time to turn to the diagrammatic conventions. First, we shall state

Types of link In an Influence Diagram

SIgnmay be omllled, bul •

Delays In Physical Flow