A statistical approach to knowledge acquisition for expert systems

Abstract

Classical expert systems are rule based, depending on predicates expressed over attributes and their values. In the process of building expert systems, the attributes and constants used to interpret their values need to be specified. Standard techniques for doing this are drawn from psychology, for instance, interviewing and protocol analysis. This paper describes a statistical approach to deriving interpreting constants for given attributes. It is also possible to suggest the need for attributes beyond those given.

The approach for selecting an interpreting constant is demonstrated by an example. The data to be fitted are first generated by selecting a representative collection of instances of the narrow decision addressed by a rule, then making a judgement for each instance, and defining an initial set of potentially explanatory attributes. A decision rule graph plots the judgements made against pairs of attributes. It reveals rules and key instances directly. It also shows when no rule is possible, thus suggesting the need for additional attributes. A study of a collection of seven rule based models shows that the attributes defined during the fitting process improved the fit of the final models to the judgements by twenty percent over models built with only the initial attributes.