THE 1987 INTERNATIONAL CONFERENCE OF THE SYSTEM DYNAMICS SOCITY. CHINA847
TOWARDS AN INTEGRATED, INTELLIGENT
KNOWLEDGE BASED SYSTEM FOR DYNAMIC
MODELLING
Dr. I. Moffatt
Department of Environmental Science
University of Stirling
Stirling FK9 4LA
Scotland, U.K.
ABSTRACT
Innovations in automated intelligent knowledge based systems (IKBS)
including expert systems (ES) could have a major impact on the development
of system dynamics methodology. This paper reports on the characteristics
of an integrated system for modelling and managing complex, dynamic systems
currently being developed. Essentially, the system consists of five
sectors namely a simulation model; a set of quantitative measures as macros;
means; .a real world data set; an adaptive knowledge based system including
an expert system; a policy decision making system. It is suggested that
such-an integrated approach to system dynamics could further enhance the
usefulness of the methodology to modelling and managing complex systems.
INTRODUCTION
Intelligent, knowledge based systems (IKBS) are part of the more general
area of computing science concerned With artificial intelligence (AI).
Studies into I.K.B.S. are concerned with, "designing computer systems that
for certain limited areas, emulate some of the characteristics of human
thought - the ability to learn, reason solve problems and understand
ordinary human language." (Shannon, et al, 1980, 276). Within this
general area of intelligent, knowledge based systems, there has been a
growing fascination with the development of expert systems (ES). One
technical approach to developing expert systems is to try and write computer
programs which will duplicate the results of learned skills or human expert-
ise without concern for whether the processes going on in a computer are
exactly the processes used by the real human expert. As Shannon puts it,
“Expert systems (ES) applications should be differentiated from pure AI
research because the primary goal of such applications is not to understand
the basic mechanisms used by the human expert to arrive at a given result,
but rather, it is to consistently duplicate the results of a human expert"
(Shannon et al, 1980, 776 emphasis added). It is against this background
of research into intelligent knowledge based systems, including expert
systems, that the rest of this paper will attempt to integrate these concepts
into conventional system dynamic modelling.
In the following section a conventional way of building system dynamic
models is described. Whilst, no doubt, there are many ways in which each
process could be described the purpose of this simple recipe for system
dynamic modelling is to illustrate the way in which system dynamic models
could be extended by integrating some concept of Al into our practice.
In section three this traditional approach to system dynamic modelling is
extetided by introducing an intelligent knowledge based system (including
an expert system) into the modelling process. It will be argued that
expert systems which work in virtual time could be very useful to model
builders and decision makers who are working with simulated and real time
respectively. The way in which an expert system can be integrated into
848THE 1987 INTERNATIONAL CONFERENCE OF THE SYSTEM DYNAMICS SOCITY. CHINA
system dynamic modelling is illustrated by a hypothetical example from
on-going research into carrying capacity. Finally, some of the unresolved
technical and ethical problems are discussed.
SYSTEM DYNAMIC MODELLING PROCEDURE
There are various procedures for building models in science (Jeffers, 1978,
Idrgensen 1986, Roberts et al, 1983). In figure 1 a tentative modelling
procedure is outlined for system dynamic modelling. As can be observed
the primary focus of all research at all times is to define the problem
clearly. This is not a trivial task as a problem that is ill-defined
is a problem which is unlikely to be solved. A problem can be defined
as an unsatisfied need to change an unusual observation to an expected
observation. A problem is solved when the unusual observation and expected
observations are perceived to be the same. Often, this procedure is not
correct at the first attempt and it needs anonstant interplay between
theoretical speculations and practical observations.
Once the problem has been clearly defined it is necessary to attempt some
form of conceptualisation of the problem viewed as a system. The second
phase of system dynamics modelling is an art which depends on the creative
imagination of the researcher as well as his/her knowledge of the way in
which the system functions. The latter can be gleaned, in part, from
detailed study in the field or laboratory, by questionnaires, statistical
analyses of data or through a thorough examination of previous studies.
In system dynamic modelling this phase of conceptualisation is achieved,
in part, by constructing a causal diagram or digraph of the system.
Model representation is the third phase of system dynamic model building.
This stage consists essentially of translating the causal diagram or digraph
into a computer flow chart. A flow chart is a partial representation
of the sequence of operations which are necessary to solve a problem.
Various conventional symbols are used in flowcharting. In system dynamics
the conventional symbols for levels, rates and auxiliary equations using
DYNAMO or DYSMAP can be inter~connected to form a multi-feed back loop
stylized flow chart. These stylized flow charts are a diagrammatic
representation of a set of completely recursive difference equations.
In the. fourth phase of model building the behaviour of the simulation
model is compared qualitatively and quantitatively with the behaviour of
the system of interest's reference mode. The reference mode represents
either an empirical trajectory of one or more state variables such as the
changing level of CO, (Carbon. dioxide) in the atmosphere or a hypothesised
mode of behaviour whfch model builders and decision makers would like a
real system to achieve such as the reduction of CO (Carbon monoxide) in
the urban environment.
Model evaluation is the fifth phase of model building and it is crucially
important that parameter sensitivity tests and careful calibration are
undertaken as well as rigorous forms of verification procedures are used
including statistical analyses. The latter can be custom built by use
of macros, Whilst the degree of correspondence between simulated and
actual data is usually measured by conventional statistical techniques it
is not the only form of model evaluation (Legasto et al, 1980). It is,
however, important to note that system dynamic modelling, as in other forms
of systems modelling, is an iterative process, It is often essential
to refine, the operational model by carefully repeating the first five
phases of model building. Often this refinement of the original model
is enhanced when quantitiative analyses are used.
THE 1987 INTERNATIONAL.CONFERENCE OF THE SYSTEM DYNAMICS SOCITY. CHINA849
The penultimate phase of system dynamic model building is concerned with
the assessment of various policy alternatives. In the original use of
the systems dynamics approach much emphasis was placed on the control and
management of industrial and urban systems (Forrester, 1961, 1969). This
emphasis on policy orientated research is still-much in evidence as
Forrester notes, "the ultimate test of a system dynamics model lies in
identifying policies that lead to improved performance of the real system"
(Forrester, 1980, 224). Despite the difficulties and ethical as well as
political implications involved in actually carrying out such tests it is
clear that many system dynamic model builders have overstressed the various
policy alternatives embedded in system dynamic models at the «expense of
more rigorous research into their own models. It should, however, be
obvious that if dynamic models are to be used as tools for implementing
environmental management or socio-economic planning then it is essential
that the model on which some of the policies may be based are sound.
If there are major weaknesses in the models thenclearly any policy
recommendations based upon them must carry little or no conviction.
(Wilson, 1970).
The final phase in system dynamic modelling is to use a well validated and
stringently tested model in order to contribute to an understanding of a
particular problem. Ina hard system this understanding may lead to an
efficient and effective solution to the problem. In soft systems, however,
this understanding may lead to political ways to promote system change and
evolution. If policies emanating from dynamic models are put into practice
then it is essential that the real world system is carefully monitored to
enquire into the ways in which the policy is effective. This does, of
course, raise important questions concerning the ideology of control
on the use of system dynamic modelling or other techniques (Gregory, 1980).
Phase 1 Problem definition <
Phase 2 System conceptualisation g
Phase 3 Model representation
Phase 4 Model behaviour
Phase 5 a evaluation 5
Phase 6 Policy alternatives
Phase 7 Model u3é@
Figure 1. System dynamics modelling: procedure.
INTELLIGENT KNOWLEDGE BASED SYSTEMS
The modelling procedure outlined above has been used in many disciplines
for several decades. Recently, however, intelligent knowledge based
systems, including expert systems, have emerged as an important area of
research activity in computing science. The term ‘knowledge base', in
this context, refers not only to the facts, but als to the relationships
and rules governing the inter-relationships between the fact and conceptual
models which have been developed in the past decade. The word 'intelligent'
vefers to the ability of the computer to make inferences from the conceptual
850 THE 1987 INTERNATIONAL CONFERENCE OF THE SYSTEM DYNAMICS SOCITY. CHINA
model and factual data; to store these inferences along with the relevant
data in its memory; and to use this new knowledge to help in the process
of solving a specific, well defined problem. These developments in
intelligent knowledge based systems have a potentially important role to
play in the future development of system dynamics modelling. One way
in which IKBS, including ES, can be integrated into system dynamics is
illustrated in figure 2. As in the earlier approach to model building
the problem is defined clearly; a conceptual model is created; a flow
chart and program are then written to create a simulation model; the model
behaviour is examined; sets of MACROS are built to calibrate and then
test the models output with the real data as part of the process of model
evaluation. At this point in the 'normal' modelling process an intelligent
knowledge based system is introduced in order to facilitate the verification
of the model and to aid decision making of alternative forecasts. As the
introduction of an intelligent knowledge based system into the normal
processes of systems dynamic modelling is relatively new it is worthwhile
explaining the way a typical knowledge based expert system can function
in system dynamic modelling. A typical intelligent knowledge based system
consists of a knowledge base, an inference engine and a workspace.
(Hawkins, 1985). The knowledge base consists of facts supplied by the
user but can also include 'facts' introduced into the knowledge base by
a simulation model acting as input. In this latter sense the inputs
from the model represent a future state of the system which, as yet, has
to be realized. The inference engine attempts to solve a problem by
searching the domain of the knowledge in the knowledge base. In the case
of an expert system the inference engine attempts to consistently duplicate
the result that a human expert would make with the same information. In
an intelligent knowledge based system this solution is stored together
with the relevant data to aid the decision making process in the future.
The workspace is an area of memory that is set aside in the computer for
the storage of the description of the problem constructed by the simulation
model or directly from facts supplied by the user. (Figure 2).
modelling procedure
(phases 1-5 in figure 1)
Intelligent, knowledge based system
Dynamic_model
B
Data base and Expert system
working space and human decision,
making space
A
Cc
Inference;
Engine
Rules hierarchy
After Pierreval and Dussauchoy,
Model Use 1986.
(phase 7)
Figure 2, Intelligent knowledge based system integrated into system dynamic modellig.
THE 1987 INTERNATIONAL CONFERENCE OF THE SYSTEM DYNAMICS SOCITY. CHINA851
In order that the intelligent knowledge based system is to be of use to
the model builder or decision makers, it is important to make a threefold
distinction in the temporal frames of reference which are used simultaneously
by people. At the first level normal real time is used by us in our
everyday activities - all our decisions are made in real time but they can
have important implications for future events in real time as well as the
other two time domains. The second time domain is that of simulated time
i.e. the length of real time simulated in a computer model. For example
in one minute of real time it is possible to simulate, say, two hundred
years of 'real' time in a dynamic model. The third temporal frame is
virtual time. This latter domain acts quickly so that the results and con-
sequencesof making one decision in the simulated time can then be examined
over the length of the simulated time horizon operating in virtual time.
This facility allows model builders to observe what would happen to the
behaviour of the model if this decision option was activated in real time
and allowed to run uninterrupted into a real future time.
The orthodox procedures used in system dynamics to evaluate forecasts are
generally to activate one or more policy option in the simulated time domain
and then observe how the system behaves. This simulated time is very small
(usually seconds for socio-economic systems) when compared to the real time
of the actual systems response to policy changes. This is shown in figure 3.
Again, the actual choice of trajectory.is made by the decision makers whilst
the model, if valid, shows a series of possible futures.
The use of intelligent knowledge based systems, including expert systems,
differs in that the future forecasts are not necessarily printed as possible
future trajectories of the system in simulated time but as possible futures
in virtual time. These possible futures in virtual time are processed
rapidly during the DT-interval of the solution to the simulated models
equations but run to the end of the length of the simulated period in the
virtual time domain. The information received by the simulated model
from these trajectories im the virtual time domain are then stored in the
workspace of the knowledge data base system. These possible futures are
then examined by the inference engine or expert system so that the 'best'
alternative including the status quo could be selected. Ina fully
automated system the 'best' option would be selected - this would, however,
assume that the expert system is as good if not better, than the frailties
of human decision makers, ~.
i | state variable
I ™~,
!
1
!
option A |- a
option B [> Yaecision Junction
eption chosen
real system
implementation
Figure 3. Virtual, simulated and real time in an expert system.
852 THE 1987 INTERNATIONAL CONFERENCE OF THE SYSTEM DYNAMICS SOCITY. CHINA
In order to illustrate the way in which this intelligent, knowledge based
system: can be integrated into normal system dynamic modelling an example
from a study of carrying capacity in Kenya will be used. (Slesser et al,
1984). In a recent study Slesser and his co-workers attempted to identify
“the consequences on national development of various actions or possible
scenarios before they are implemented in order to see whether they. are going
to create a)more sustainable or more developed society, given the existing
population's resources, given the socio-cultural pattern and given its
present state of development" (Slesser et al 1984, p.18). By building
a simple system dynamics model named ECCO (Enhanced Carrying Company Options)
a whole series of possible projections of development were produced.
These scenarios ranged from unsustainable forecasts to several options
which, in theory, are sustainable. By embedding the expert system with
the structure of this model it is possible for the IKBS to select the
"best' option at the current real and simulated time. Furthermore, it
is algo possible for the. expert system to examine possible scenarios in
+irtudl time to see what would happen if the apparent 'best' scenario was,
effected by a stochastic disturbance. Further work into this interface
is under-way.
What advantage, if any, does this new approach to system dynamic modelling
have over the orthodox procedures? The major advantage is that the use
of an expert system in system dynamic modelling integrated in an adaptive
knowledge based system permits the computer to select the 'best' option
from its examination of the variety of options simulated in virtual time.
Normally, policy makers only select the most suitable action from an
examination of the trajectories of several forecasts printed out as
simulated futures. Whilst there is little wrong with this procedure the
major difficulty arises when one option is chosen from the simulation model
and implemented in a real time system. Once the decision maker has chosen
this course of action it is often difficult to reverse that decision and
choose another more appropriate form of action. By simulating the
options in virtual time and using the knowledge gained from these scenarios
(stored in the data base) together with some objective choice directed by
the expert system it is possible to quickly ascertain the ‘best’ possible
policy to guide the real world system.
There are, however, three important problems which still have yet to be
resolved. The first technical problem requires that the simulation in
virtual time is completed very quickly - withinthe solution time (DT) of
the model's simulated time - before proceeding with the dynamic simulation,
This requires the use of efficient parallel processors which can explore
a whole variety of options simultaneously before permitting the main DYNAMO
program to proceed} The development in both hardware and software should
eradicate or, at least, reduce this problem. A second technical problem
is to overestimate the reliability of the 'expert system’. Despite the
rapid advances in this area of research into artificial intelligence the
expert systems are still crude. As Hawkins notes the user must recognise
"the limits of the domain of expertise of the system. Expert systems can
be useful; they can also be hazardous....The most obvious danger would
be for a user to rely on.a program that wasnottotally expert; one that had
significant deficiencies in-its knowledge base ... " (Hawkins, 1985 p.16).
Third, even when thesé technical problems are overcome it is important that
the models are robust and carefully tested before policies emanating from
these models are put into practice. As Hare comments: "it is the public's
privilege to choose politically among the options, but the scientist has
to specify. the range of possibilities and the means whereby the goals may
be reached." (Hare, 1983, p.136). Clearly, the development of an
integrated, intelligent knowledge based system for. dynamic modelling is
still in its infancy.
KT
THE 1987 INTERNATIONAL CONFERENCE OF THE SYSTEM DYNAMICS SOCITY. CHINA853
CONCLUSION ,
This paper has reported on the way in which expert systems can be embedded
into the normal system dynamics modelling paradigm. Unlike the normal
system dynamics paradigm, however, some forms of the decision making options
for alternative policy evaluation are undertaken by the expert system rather
than by the human decision maker. The way in which this technical problem
is resolved is by allowing the DYNAMO compiler to work in virtual time to
discover the possible impacts of these decisionsif the latter were
implemented in the real world. By using a form of time warping in system
dynamics a kind of relativity dynamics is developed (Keloharju, 1983;
Zhixin et al, 1986). An illustration of this approach was. given in the
case of modelling carrying capacity options for a country.
The major advantage of this system.is that during each simulation the
expert system can remember the impact of previous simulated runs in virtual
time and then decide which is the 'best' option for the simulation to
proceed and calculate actions to ensur.. The major disadvantages in this
approach is that it is technically difficult to allow time warping using
DYNAMO. Furthermore, neither the expert system nor human decision makers
are perfect - the best policy run may not meet the approval of other
actors in the real system (Gardiner and Ford, 1980). Nevertheless, this
integrated approach to system dynamics once developed has the potential
for enhancing the methodology for modelling and managing complex systems.
REFERENCES
Forrester, J.W. (1961) Industrial dynamics MIT Press, Cambridge, Mass.
Forrester, J.W. (1969) Urban dynamics MIT Press, Cambridge, Mass.
Forrester, J.W. and Senge, P.M. (1980) Tests for building confidence in
system dynamic models. In Legasto, A.A.. et al, ibid.
Gardiner, P.C. and Ford, A. (1980) -Which policy is best, and who says so?
In Legasto, A.A. et al, ibid.
Gregory, D. (1980) The ideology of control: systems theory and geography.
T.E.S.G.. Ixxxi, 6, pp.327-62.
Hare, K. (1985) Future environments: can they. be predicted., T.1.B.G.
10, 2, pp.131-137. oat
Hawkins, R.D. (1985) Artificial intelligen¢e from the systems engineer's
viewpoint. In Holmes, W.M, (eds) ibid.
Holmes, W.M. (Ed.). (1985) Artificial Intelligence and Simulation.
SCS Publication, California.
IJgrgensen, S.E. (1986) Fundamentals. of Ecological Modelling. Elsevier
Publications, Amsterdam, ue
Keloharju, R. (1983) Relativity dynamics, Acta Academiae Oeconomicae
Helsingiensis, Helsinki. Shae
Legasto, A.A., Forrester, J.W. and Lyneia, J.M. (eds.), (1980) System-
dynamics Vcl.i4, TINS Studies in Menegement, Amsterdam.
854 THE 1987 INTERNATIONAL CONFERENCE OF THE SYSTEM DYNAMICS SOCITY. CHINA
Mouris, E. and Sopers, F.P.M. (1986) Modelling aspects of the sugar industry
and simulation aspects of manufacturing and energy. In Vansteenkiste,
G.C. et al., ibid. pp.684-688.
Pierreval, H. and Dussauchoy, A. (1986) Use of expert systems to complete
traditional approaches for production systems simulation modelling.
In Vansteenkiste, G.C. et al, ibid. pp.377-383.
Roberts, N., Andersen, D., Deal, R., Garet, M. and Shaffer, W. (1983)
Introduction to computer simulatior the system dynamics approach
Addison-Wesley Publishing Company, Reading, Mass.
Shannon, R.E., Mayer, R., Adelsberger, H.H. (1986) Expert’ systems and
simulation. Simulation, June, pp.275-284.
Slesser, M. and Hounam, I. (1984) Carrying capacity assessment with a
pilot study of Kenya. U.N.E.S.C.0. Final Report, Energy Studies Unit,
Strathclyde University.
Vansteenkiste, G.C., Kerckhoffs, E.J.H., Dekker, L. and Zuidervaart, J.C.
(Eds.) (1986) Proceedings of the 2nd European Simulation Congress,
Antwerp, Belgium.
Wilson, A.G. (1970) Book review of Urban Dynamics by Jay W. Forrester.
Environment and Planning, 2, 3, pp.364-65.
Zhixin, D., Heyins, Z., Romous, G. and Unger, B. (1986) Distributed
systems in Simulation: data base. concurrency control using time warp.
In Vansteenkiste, G.C. et al. op.cit. pp.17-19.
