Reference: Shortliffe, E. H. AI Meets Decision Science: Emerging Synergies for Decision Support. Springer-Verlag, Berlin, 1991.
Abstract: In the 1970s, the field of medicine forced clinically oriented AI researchers to develop ways to manage explicit statements of uncertainty in expert systems. Classical probability theory was considered and discussed, but it tended to be abandoned because of complexities that limited its use. Furthermore, many AI researchers were fundamentally disinterested in normative probabilistic models because such approaches were viewed as having no relationship to actual human problem-solving methods and were a distraction from the study of the nature of expertise and the way in which human beings reason under uncertainty. In medical AI systems, uncertainty was handled by a variety of ad hoc models that simulated probabilistic considerations. Largely ignored, despite the need for some of the systems to suggest therapy or other management plans, were explicit value models. Instead, the ad hoc models were imprecise meldings of probabilistic and utility notions. This paper summarizes recent work in our laboratory which has attempted to show that formal normative models based on probability and decision theory can be practically melded with AI methods to deliver effective advisory tools. There has now been substantial work illustrating the synergies between formal decision science and AI. Newer graphical interactive techniques, and the increasing computing power of modern machines, have made decision-theoretic approaches to uncertainty management practical in a way that they would not have been in the 1970s. I am increasingly persuaded that ad hoc approaches to uncertainty management in medical advice systems can be abandoned in favor of more rigorous and formal, axiomatically derived techniques.
Notes: Revised February 1993.