Graham W. Winch
Professor of Business Analysis
Plymouth Business School
Drake Circus
Plymouth
PL4 8AA
England
It was once suggested to me, half jokingly perhaps, that System
Dynamics offers "2nd-Generation Expert Systems - before the ist
Generation". This paper reconciles the theories and processes, and
draws upon business consulting assignments, to examine how close to
reality this notion is.
© what are Expert Systems supposed to do?
A very basic description has been offered by Clifford et al (1986):
"An expert system is a computer application that provides decision
support similar to that of a human expert in solving problems".
or similarly (Hertz, 1988):
".... computer programmes that provide advice and diagnoses for
advising problems ordinarily dealt with by human experts".
Most expert systems that have been developed to date are in scientific
fields, for example, medicine, engineering and geology. Conversely it
has been asserted (Coombs & Atty, 1984) that despite significant effort
by the Artificial Intelligence community far fewer substantial
294
applications have resulted than might be expected - particularly in the
business decision field. Cullen and Bryman (1988) in a survey of 70
expert systems concluded that successful applications are. most likely
in:
- narrow rather than wide knowledge domains.
- where required expertise is mainly factual rather than
procedural.
= systems contain shallow knowledge rather than deep underlying
knowledge.
This points to a technology which may have great potential, but with
the exception of a few, mainly factually-based scientific systems, has
failed to provide many fully fledged working systems which offer true
human-like knowledge and reasoning.
e Are system dynamics models expert systems?
In basic terms expert systems are comprised of the following elements:
1. factual and procedural information,
2. rules that describe the relationship between stored data and
inputs,
3. methods, programmed within the system, for coming to conclusions
based on the stored "knowledge", and
4. an interface with the user for receiving his/her queries, offering
results and for explaining the need for inputs and the system's
reasoning.
Waterman (1986) has identified key features of expert systems by which
he differentiates them from conventional (sic) systems:
295
EXPERT
SYSTEM
Expertise [— Exhibit expert performance
Have high level of skill
— Have adequate robustness
Symbolic --— Reformulate knowledge symbolically
Reasoning '— Reformulate symbolic knowledge
Depth [_ Handle difficult problem domains
Use complex rules
Self Examine own reasoning
Knowledge L Explain its operation
On first glance the purpose of a business expert system could be the
same as for a system dynamics model, and indeed the processes in
model/system development may be very similar (Winch, 1984). However on
strict definitions a typical system dynamics model could not be classed
as an expert system, as it will inevitably lack a number of these
features. That said, many so-called expert systems also lack a number
of these features, and have been criticised for brittleness (Hudlicka,
1988), failure to exhibit true expert performance - "overblown accounts
of what they can do" (Towris, 1986), and (as discussed earlier) a low
presence in difficult problem domains with complex rules.
However, system dynamics models, particularly the large scale business
analysis and forecasting models with which the author is familiar, do
encompass many of these features. They manifestly contain information
concerning facts, identities and procedures within the modelled
systems, and rules or inter-relationships that link and inter-link that
information with inputs. Indeed recent ES literature has begun to
296
advocate and utilise causal relationships as opposed to "rules-of-thumb
gleaned from experience", with the assertion that causal models,
because they represent generalised understanding rather than specific
problem-solving rules, are likely to be more versatile (Basden,
1984).
Hudlicka (1988), using arguments familiar to system dynamics
practitioners, asserts that in constructing causal models for expert
systems, ES practitioners are forced to understand the problem domain
well enough to express formally its structure and behaviour. He
continues that this type of understanding is qualitatively different
from that required when knowledge is represented by uncorrelated rules,
"as was the case with the first generation of expert systems".
This freedom has enabled the system dynamics modeller to adapt and use
generic structures within different business models; for example, plant
start-up and shutdown decisions, capital investment decision-making,
product and technology upgrading. Apart from the obvious value of "one
model many uses" economics, this bring two further benefits:
- these generic structures are continually being verified and
modified, representing a learning process.
om a developing model typically combines such existing knowledge with
different domain expertise for the new application.
This second point is particularly important in terms of the "self
knowledge" feature in Watermans' earlier definition. Although system
dynamics models do not generally permit interrogation by the user of
the system for explanation of reasoning or input requirements, the
comparatively transparent code in DYNAMO, STELLA and other similar
syntax languages means that the nature and requirement for particular
information can be readily reviewed by the user in both existing
generic structures as well ‘as in the developing new domain areas. This
self-knowledge aspect is enhanced by causal-loop or influence diagrams,
and obviously particularly so by the on-screen flow diagrams with
STELLA which vividly portray the knowledge of inter-relationships
captured in the model.
297
DYNAMO, along with other packages with similar syntax, is a non-
procedural. language. It therefore shares the model building and
implementation advantages over procedural languages with declarative
languages like PROLOG, which Hampshire (1988) contends are ideal for
simulating real-world processes. (Though, even with. a declarative
language like PROLOG, some find that adding new rules as new knowledge
is acquired may need restructuring of the programme (e.g., Clifford et
al., 1986) - a rare requirement with DYNAMO). In fact, Hampshire
maintains that an expert system is really just another form of
declarative language, with only the particular distinction of being
able to explain why it comes to a conclusion. It has to be
questionable if with current hardware and software any internal
reasoning and explanation facility could be built into an expert system
that could (realistically) explain loop-gain in typical system dynamics
models, or the complex inter-relationships and frequently observed
counter-intuitive behaviour of such complex systems. The combination
of causal-loop diagrams, transparent code, and the system dynamics
modeller does at least provide, currently, a satisfactory hybrid system
for achieving this.
In one final, interesting respect system dynamics models may be similar
to expert systems. One way that Artificial Intelligence approaches are
argued to be able to reflect real-world processes is in dealing with
uncertainty, particularly by using fuzzy logic (Zadeh, 1979). In this
concept, rules or facts are not defined with binary precision, but
rather values or distributions are. assigned to indicate the degree or
extent to which a description or factor applies, e.g., instead of
defining tall as over 1.70 m (thereby consigning those 1.70 m or under
to be not-tall or short), fuzzy logic permits heights to be assigned
tallness quotients, representing the extent to which they conform to
the adjective "tall" - 1.7 m might be 0.7,. whereas 2.00 m would be more
like 0.95. | Systems using this can combine such uncertain or . fuzzy
variables and relationships. into their reasoning.
However system dynamics modellers are well used to applying multipliers
from table functions to represent similar fuzzy concepts in their
models. For example, a decision to expand plant capacity may be based
on current conditions concerning expected profitability of the product,
projected growth in demand, closeness of current plant utilization to
298
desired maximum, availability of finance. Variables representing each
of these factors can be converted into a quotient or multiplier and
combined to give an overall propensity or desirability to expand which
would control an expansion decision (probably through an intension-to-
build backlog that would trigger at an expansion increment).
e What is the relationship between ES & SD?
In a number of important respects system dynamics models involve many
of the essential elements of expert systems, more if the whole process
of modelling is included:
= SD models comprise formulations capturing knowledge about the
structure and decision-making processes of real-world systems.
- They are able to cope with deep and underlying knowledge as well
as simple rules, and are generally robust.
- In terms of the comparatively transparent code of DYNAMO syntax
and of causal-loop, influence and flow diagrams there is an
ability to explain to users how results occur, behaviour is
generated, and inputs/parameters fit in.
= If the “system" is allowed to include a system dynamics
practitioner then he can help elicit system behaviour and
understanding - even though he may not be an expert in the
particular domain.
Indeed, in as far as many claimed expert systems do not encompass all
ef these either - they may only be simple rule chains, may deal with
only trivial or superficial representations, may not have-complete
explanations of reasoning etc., - it could be argued that at least in
terms of functionality, a well designed and implemented system dynamics
model could fit the bill just as well. It would probably also be much
cheaper to implement for any large (complex) real-world systen,
especially for systems of the "practitioner assistant" rather than
"expert consultant" type.
299
In the future this "competition" will become irrelevant. Already it
has been recognised (Basden, 1984) that there needs to be a “blurring
of the distinction between ‘expert systems' and ‘conventional’
computing techniques, so that techniques are used according to their
usefulness rather than their label". A hybrid model has already been
reported by Levary & Lin (1988) called HESS which embeds a
software lifecycle simulation model written in DYNAMO within a system
which includes an input expert system to check compatibility of input
vector, an output expert system which makes recommendations regarding
the software development process, and a knowledge base management
system which logs and reconciles input vectors with recommendations.
Future generations of computers, particularly parallel processors,
offer the likelihood of full integration between these technologies.
Basden A. (1984) "On the Application of Expert Systems" in Developments
in_Expert Systems, ed M. J. Coombs, Academic Press.
Clifford J., Jarke M., Lucas H. Cc. (1986)
"Designing Expert eyetene ina _Business Environment" in
Art: rel ite: e nagement, ed. L.
F. Pau, Elsevier science Publishers BV (North-Holland) .
Coombs M. & Atty J. (1984). “Expert Systems: An Alternate Paradigm"
in Developments _in Expert Systems, ed. M. J. Coombs, Academic
Press.
Cullen J. & Bryman A. (1988). "The Knowledge Acquisition Bottleneck:
Time for re-assessment", Expert Systems, Aug., Vol 5, No. 3.
Hampshire N. (1988). "Introducing Declarative Languages". EXE
Magazine, October, Vol 3, Issue 5.
Hertz D. B. (1988). The Expert Executive, John Wiley & Son.
Hudlicka E. (1988). “Construction and Use of a Causal Model for
Diagnosis." Int. J. of Intelligent Systems, Vol 3, No. 3.
Levary R. R. & Lin, Chi Y (1988). “Hybrid Expert Simulation System
(HESS)", Expert Systems, May, Vol. 5, No. 2.
Towris J. “Expert systems in Freight Management." SERC Workshop on
Expert Systems in Transport, Univ. of Leeds/SERC.
Waterman D. A. (1986). A _Guide to Expert Systems, Addison-Wesley.
Winch G. W. (1984). "The Simulation Process and Techniques" in Policy
Evaluation using REVEAL, ed. M. small, IcL
Publications.
Zadeh L.A. (1979). "A Theory of Approximate Reasoning". Machine
Intelligence, Vol 9.
