Heuristic Abstraction in the Decision-Theoretic

A criticism of diagnostic systems, which are based on the formal foundations of probability and utility, is that their reasoning strategies and recommendations are in°exible and unnatural. We have developed a facility that increases the °exibility of normative reasoning systems by providing multiple human-oriented perspectives on diagnostic problem solving. The method endows a system with the ability to reason about arbitrary classes of diagnostic entities and to control the level of abstraction at which inference occurs. The techniques have been integrated into Pathflnder, an expert system that performs hematopathology diagnosis. We explain the background and approach that we have taken, and describe how we use the techniques in Pathflnder to modulate information- and decision-theoretic reasoning with strategic scripts that are familiar to physicians