A PROBABILISTIC THEORY OF MODEL-BASED DIAGNOSIS

Diagnosis of a malfunctioning physical system is the task of identifyi ng those component parts whose failures are responsible for discrepanc ies between observed and correct system behavior. The result of diagno sis is to enable system repair by replacement of failed parts. The mod el-based approach to diagnosis has emerged as a strong alternative to both symptom-based and fault-model-based approaches. Hypothesis genera tion and hypothesis discrimination (action selection) are two major su btasks of model-based diagnosis. Hypothesis generation has been partia lly resolved by symbolic reasoning using a subjective notion of parsim ony such as non-redundancy. Action selection has only been studied for special cases, e.g. probes with equal cost. Little formal work has be en done on repair selection and verification.This paper presents a pro babilistic theory for model-based diagnosis. An objective measure is u sed to rank hypotheses, viz., posterior probabilities, instead of subj ective parsimony. Fault hypotheses are generated in decreasing probabi lity order. The theory provides an estimate of the expected diagnosis cost of an action. The result of the minimal cost action is used to ad just hypothesis probabilities and to select further actions. The major contributions of this paper are the incorporation of probabilistic re asoning into model-based diagnosis and the integration of repair as pa rt of diagnosis. The integration of diagnosis and repair makes it poss ible to troubleshoot failures effectively in complex systems.