A System for Empirical Experimentation with Expert Knowledge

When CASNET (Chapter 7) evolved into the general s,vstem-building tool known as EXPERT, one of the first applications was a rheumatology consultant program called AI/Rheum (Kingsland and Lindberg, 1983). Developed collaboratively by researchers at Rutgers University and the University of Missouri, AI/Rheum quickly became large and complex, thereby complicating the process of knowledge base maintenance. Peter Politakis, a Rutgers graduate student working with Sholom Weiss and Casimir Kulikowski, accordingly developed a program, named SEEK, that was designed to assist with both expansion and ver~fication of the AI I Rheum knowledge base. SEEK illustrates how a model ~f expert reasoning (in this case the rules of rheumatology diagnosis) can be rf:fined with program assistance. The program suggests possible experiments involving generalization or specialization of the preexisting rules in the system. A library of stored patient cases with known conclusions is used as a basis for proposing the experiments. This approach has proven particularly valuable in assisting the expert in a domain like rheumatology where two diagnoses are often difficult to distinguish. The research on SEEK also has its origins in the knowledge-acquisition tool TEIRESIAS, developed by Davis for lHYCIN (Davis, 1979). However, SEEK is able to go a step further by using a somewhat more articulated representation than MYCIN's rules. In AI/Rheum evidence is class~fied according to major and minor findings, plus required and