[Hosts Note : The discussion on optimization is beginning to drift a
little from issues in system dynamics. Nonetheless, the
following comments, I think, bring forth some very important
issues in system dynamics.
Jay Forrester has, for decades, advocated the use of models
to develop policies that will improve performance of social
systems. At the same time the words "optimal" and "best" are
typically (though not universally) downplayed. But certainly
Forrester never advocated the use of bad policies. If we replace
"optimial" with "really great" do any objections remain. Now the
question becomes how formal a model do you need to develop a
really great policy?
The enclosed note talks about optimizing informal qualitative
models. The question this raises is whether this is good
enough. Or, turned around, isnt that how Bosnia got to be
the way it is? ]
The concept of optimizing a systems behaviour, I believe, is independent
of the the domain under investigation. What is more important is the
nature of the model, the model attributes, and the optimization goals.
Engineering applications are usually concerned with quantifiable processes
(almost by definition), thus the models are numeric and the optimization
occurs over a multidimensional solution surface represented by numerical
values.
A simple example of a qualitative system: an interior decorator optimizes
over physical layout, light, architectural features, color balances,
and personal preferences. Interior decoration is not, as yet, a capability
easily transferred to a computer model, but it would seem to me that there
is the essence of this type of problem is an interesting research topic, very
relevant to system dynamics.
Optimization does have have a social relevance, at least local optimization,
as opposed to global optimization, but the tools are different. Perhaps the
tools are not even analytic, but that does not mean an optimal solution
in some sense is not derivable. The important issue is the derivation of
models for social systems that can be operated upon systematically and
deterministically to produce a relevant, practical optimal solution.
It is often assumed that quantitative representation is required for
implementation of the model and the optimization methdology in a computational
form that can be programmed and executed in a computational form. Most
expert system models are implemented in this manner: subjective or qualtitative
information is mapped into a quantitative representation to allow computational
operations. Thus, to continue the example, the compatibility of certain color
schemes might be mapped to a numerical range of 0 (green with red) to 1
(burgandy and beige).
Because our current computational models are quantitative (i.e. numeric), we
tend to think of optimization as being numeric. Until we develope automata
that can operate on relational models (neural nets may be a basis for such
a solution) the need to instantiate a model in computational terms should be
divorced from the modelling domain itself.
--
Gary W. Kenward
ITO HEWLETT-PACKARD (Canada) Ltd. email:
gary@idacom.hp.com
Suite 2301 phone: 604 454 3439
4710 Kingsway Avenue fax: 604 454 3401
Burnaby, BC Canada V5H 4M2
"To see a world in a grain of sand,
And heaven in a wild flower,
To hold infinity in the palm of your hand,
And eternity in an hour." -- William Blake