John Mark Agosta

Room 501, IS Building
Tuesday, November 15, 2005
Welcome Coffee - 10:30-11:00 am
Talk - 11:00 -12:00Noon

Abstract: A technician's commonly held view is that diagnostic tools should assemble diagnostic knowledge "automatically" from successful past diagnoses, so that diagnoses can be replayed when similar symptoms arise. This contrasts with probabilistic knowledge-based systems methods where significant time and effort is needed up-front to create diagnostic models. The naive solution considers diagnostic knowledge as a database of cases, composed of a set of symptoms that explain the underlying failure, and the process of diagnosis as one that consists of database lookup. Of course this retrieval model for diagnosis is a weak representation compared to a Bayes network representations. The question is, can we have the best of both worlds---to satisfy the technician's desires while exploiting the elegance and power of Bayes networks diagnostics?

If we start from this assumption, the challenge is if we can construct a Bayes network consistent with the known set of cases. Unlike a statistical learning paradigm, each case is considered a probabilistic constraint to which the convergence of the ultimate model is subject, rather than a noisy realization of a partial instantiation of the variables.

This talk will specify this problem precisely, and suggest how it to approach it.

Speaker's Bio:John Mark Agosta develops diagnostic, optimization and statistical models to improve Intel products and processes. Previously he has worked for Edify Corporation, Knowledge Industries, and at SRI International, building diagnostic and planning models for medical, equipment, and emergency response applications. Agosta received his Ph.D. in the Stanford University's Engineering-Economic Systems Department (now Management Science and Engineering) in 1991.