Thanks to all the list members for their many helpful references. Almost all of these pointed me towards literature on the use of temporal Bayes nets. I am investigating these, but the question I was trying to ask, but not clear enough about was the following: In conventional, non-temporal diagnosis, how do we come up with acceptable networks. Here's a made-up example of what I want: Consider a hypothetical (implausible) immune system disease, D. D has the strange effect of striking its victims at the extremities first, then progressing inwards. A patient presenting in the early stages of the disease might present with symptoms like "infected hangnail," and "gangrenous toe." A patient presenting in late stages of the disease, though, would be like a late-stage AIDS patient, riddled with opportunistic infections. I could imagine a (non-temporal) Bayes net for this diagnostic problem having two disease hypotheses: "D early stage" and "D late stage." "D early stage" would cause "infected hangnail," and "gangrenous toe." "D late stage" would cause "infected hangnail," and "gangrenous toe," but also "pickled liver" and "watery lungs" (via "pneumonia"). My question is, has there been any thought about principled ways to choose how to introduce hypotheses like these? How do we know that we need two diagnostic hypotheses for D, or three or four, as opposed to 27? (Perhaps in medicine this is just part of the empirical wisdom of the field.) Also, has there been any thought about incorporating "late stage" disease diagnoses with other hypotheses in a single diagnostic engine? I.e., having a hypothesis that effectively can explain ANY observable symptom? The issue relates to my original question of how to diagnose failures of devices that have feedback, since if a failure persists for some time in such a system, that system will start displaying "late stage" symptoms --- the tight coupling of the system will cause malfunctions to appear everywhere. Best, Robert
