Use of Elvira’s explanation facility for debugging probabilistic expert systems

Debugging an expert system is virtually unfeasible without explanation facilities, especially in the case of probabilistic expert systems, whose way of reasoning is completely different from that of human experts. Unfortunately, almost currently available tools for building probabilistic graphical models offer no explanation facility. This paper shows how the explanation capabilities provided by Elvira, a software tool for editing and evaluating probabilistic graphical models, have helped us in the debugging of two medical Bayesian networks: Prostanet, for the diagnosis of prostate cancer, and hepar ii, for liver disorders.