Certainty factor theory and its implementation in a medical expert system shell.

Abstract

In this article some topics related to the MYCIN-model for uncertainty handling are discussed. First, it is pointed out that one of the problems identified by Adams should be considered as evidence supporting the point of view that the probabilistic interpretations of measures of belief and measures of disbelief given by Shortliffe are inappropriate. In this context some results achieved by Heckerman are reviewed, and it is shown that similar problems can be identified in the interpretation given by him. It is also pointed out that Heckerman's argument that the original probabilistic interpretation of certainty factors implies non-commutative evidence-updating is not correct. Then, in the context of the result achieved by Horvitz and Heckerman, in which it was proven that there is an isomorphic mapping from the evidence combination scheme for the likelihood ratio to the evidence combination function in EMYCIN and its successors, it is pointed out that the evidence combination scheme for the likelihood ratio suffers from a great restriction on its applicability in expert systems. Subsequently, some short remarks are made with respect to the complaint that sometimes the MYCIN model cannot work as well as expected. Lastly implementation of the certainty factor mechanism in a HIS-oriented expert system shell with some substantial changes is described.