This is a very interesting and useful discussion. I have a few
points/clarifications to add:
1. I agree structure of the model is very important - but if its output is
not checked with real data, then this structure is as limited as the mental
model that produced it. I dont know if any research has been conducted on
this issue, but to me it seems that, given, exactly the same problem,
different system dynamics modelers could come up with different structures
to explain it. There is no way to guarantee/predict what will happen in
future, but we can have more confidence if the model explains a) relevant
historical behavior (so data-fitting/automated calibration here is
important); b) relationships used in model are good (this is where
economists criticize SD; hence again data-fitting/automated calibration
here is important). Assuming that there are different SD models
(structures) using similar historical data but different structures, and
have similar outputs, then Id choose the model structure (as
physicists/mathematicians choose amongst competing theories) that is
simplest and most elegant.
Creating a model with fuzzy, untested relationships and/or fuzzy data is
okay too, at least to start with, because simulations can point us to the
need for better data
elationships/structure.
2. Data is very important too, as I mentioned in my post - the story I
remember is that Lorenz was working on the weather model, represented by
the now-famous three differential equations (the Lorenz system). Late
night, graduate student life - he makes a small change in one of the
parameters, goes out for a coffee, and comes back expecting nothing
earth-shattering. But there on the screen the whole pattern represented by
the state variables was completely changed and was totally unexpected
(emergence of chaos!!) - and the rest is history. So accurate data can lead
to new insights too. By the way, this was a serendipitious discovery and
the parameter choice was neither based on model nor real data. As our
knowledge of nonlinear mathematics grows, hopefully we will have better
bounds on the parameters to test for chaotic behavior.
Bottom line repeat: both are important, sometimes structure gives more
insights, sometimes data does.
3. Thanks for your clarifications Alexandre on differentiating between
calibration and optimization. Benny mentioned that the structure in reality
is not fixed, so "optimal control" is not really optimal, and automation is
not the best way to get to it. I agree that the structures
elationships
change in reality, but I still maintain that search for optimality, for a
given fixed structure, is best done by automation. Here is how I think of
this:
What is the best control if structure is S1, data is D1
What is the best control if structure is S2, data is D1
What is the best control if structure is S3, data is D1
....
If data is fuzzy, D1 could be replaced by D2, D3, etc.. This "what is the
best if" is in contrast to the usual "what-if" that is done in SD. Search
for the "best control", for each of the above, is best done by automation.
The knowledge that results from repeated optimal search would be an order
of magnitude more (practically) useful than that obtained by
simulation-based policy analysis alone (call the above optimlations or
simulzations;-)), and they will get more useful as the
data
elationships/structure defining the models improve. Why do we need
them? Because we are dealing with large and complex systems and just as
mental models are improved by SD, our search for better controls is
enhanced and bounded by (repeated) optimizations. [Stochastic optimal
control/dynamic games may be another, direct but more difficult, way to
deal with uncertain and changing realities]. Of course, as I have mentioned
b4 on the list, if the models are not dependable, optimization is meaningless.
I just think this is the current trend anyway - as computers and software
become more powerful, optimization would be the way to get better
solutions/productivity/profits etc. (Btw, Optimal control may not be the
best way for this optimization either - genetic algorithms/AI techniques
may be used, but I dont know much of them, tho recently saw an
interesting appln from Brigham Young Univ, applied to infrastructure systems).
Thank you very much for your time
Jaideep
From:
jm62004@Jetson.UH.EDU
****************************************
Jaideep Mukherjee, Ph. D.
Research Associate
Department of Industrial Engineering
University of Houston
4800 Calhoun Road
Houston, TX 77204-4812
Phone: 713 743 4181; Fax: 713 743 4190
****************************************