Performance evaluation of an expert system using rescaled certainty factors (medical diagnosis)

Certainty factors are frequently used as an alternative to probabilities for representing uncertainty in expert systems. Rules proposed for combining certainty factors are much more easily applied than are the Bayesian rules for revising probabilities, since Bayesian rules require sequential assessment of conditional likelihood functions. However, an inconsistency arises in the results obtained when certainty factor rules are applied in sequential steps to update certainty factors having differing signs. A possible simple solution to this problem, involving rescaling the certainty factors, is discussed.<>