I don't know of much in the literature on the general problem of using
Bayes nets to diagnose systems with substantial feedback aspects
(including the human body!).
Most of the work on medical diagnosis that I have seen seems to assume
a diagnostic model where the evolution of diseases is factored out.
I.e., there's a well-understood set of root level diagnoses and the
pathologies and symptoms are distinguished from these, so we don't
worry about feedback.
This approach seems well-suited to medical diagnosis, where there is a
substantial body of knowledge about how patients actually present, but
what about diagnosing devices with feedback? Presumably one needs to
be able to synthesize a reasonable model of the state in which the
device will present.
I believe Greg Provan did some work on this problem in medical
diagnosis, using temporal influence diagrams. Can anyone point me to
other work in this area?
Thanks,
Robert
