Reappraisal of the use of conditional probability in early expert systems

摘要

A qualitative method of reasoning under uncertainty using Bayesian probability was originally proposed by Duda et al. However, this has been found to be 80% reliable at best in practice, and better techniques are being looked for by those who seek greater accuracy. The paper concludes that Bayesian approaches should not be abandoned. A Bayesian approach can be used as a sound quantitative technique, provided that the axioms of probability are properly observed and the nature of Bayes' theorem itself is understood. The paper documents some lessons that can be learned from the previous use of Bayes' theorem, and assists in the understanding of the way that it works in practice. The limitations of assuming conditional independence are emphasized, and the consequent equal ratios that must exist between various pairs of joint probabilities are highlighted. This leads to validity checks that can be made by an expert-system designer using local calculations and Bayesian update methods. It is also demonstrated that, for even modest numbers of antecedents, the assumption that these are independent overwhelms the information given by the expert, who, when considering a particular piece of evidence, has control of little more than the ‘all other antecedents true’ and the ‘all other antecedents false’ scenarios. The same conclusions may be pertinent to more recent work, but this is not explored in depth.