I have lately been challenged by the following problem: Suppose
a model of some system is expressed as a computer program. It seems
possible, in principle, to analyze the program to generate a data-
flow graph and then use that to construct a Bayesian network. The
Bayesian network could be used to compute the same result as the
program, but it could extend the results generated by the program
by taking uncertainty in parameters into account and by allowing
data to flow from bottom to top as well as the usual top to bottom.
The particular programs I have in mind are models of heating, venti-
lating, and air conditioning equipment which are based on physical
principles. Quite a few such models have been constructed, and these
represent a very sizable database of engineering knowledge. Translating
this information into probabilistic models would be greatly speed the
construction of practically useful Bayesian networks.
I haven't been able to locate a reference to automatic construction
of Bayesian networks -- perhaps you can point me to one. I don't
know much about graphical program analysis, so if you have a
reference for that, I would be very interested.
Regards,
Robert Dodier