Diagnosis based on incomplete causal model

Abstract

A mechanism for diagnosis of multiple faults in a system has been implemented. The domain knowledge is represented at the deep level by a causal network and at a shallow level by rules. Two forms of incompleteness are accommodated in the model: nonverifiability of initial causes and incompleteness in the relations in the causal network. The diagnostic process is non-monotonic and may admit multiple solutions. It proceeds in iterations of set covers over the set of all observations. The criterion of parsimony over initial causes has been used while reasoning at the deep level to generate a 'IC-Parsimonious' solution which ensures minimal number of initial causes. The system then abstracts shallow knowledge from this in the form of rules. When a diagnostic problem is posed, the system tries to generate a solution using the shallow knowledge. If it does not succeed, then it uses the deep knowledge to generate a solution. >