edith@ai.toronto.edu (Edith Fraser) (03/01/90)
Department of Computer Science, University of Toronto (GB = Galbraith Building, 35 St. George Street) ------------------------------------------------------------- ARTIFICIAL INTELLIGENCE GB305, at 11:00 a.m., 15 March 1990 Scott Goodwin University of Alberta "Statistically Motivated Defaults" A fundamental area of research in Artificial Intelligence is the development of a computational theory of common sense reasoning. My work, as well as the work of several other researchers, has shown that the foundations of such a theory can be laid by viewing common sense reasoning as a rudimentary kind of scientific reasoning. This view is solidly rooted in the philosophy of science. My colleagues and I have proposed a simple hypothetical reasoning framework, called THEORIST. We have shown the conceptual efficiency of this framework and its applicability to a wide range of common sense reasoning tasks such as reasoning by analogy, planning, diagnosis, and image understanding. My work focused on one aspect of THEORIST, viz., the semantics of defaults. In common sense reasoning, some knowledge is generally true but admits exceptions (e.g., birds fly, objects retain their colour when moved, and people with colds cough). Knowledge of this type is referred to as defaults. I view a default as a possible hypothesis that can be assumed in some cases. Further, I view a default as making a statistical claim about the world. It is the statistical knowledge encoded by the default that justifies assuming it in particular cases. In general, we may know many things about particular cases but not all of this knowledge is relevant to the applicability of defaults. In probability theory, the problem of identifying the relevant knowledge is called the problem of choosing the reference class. One contribution of my work is in addressing this problem. Additional problems arise in reasoning about time. In particular, there is the problem of persistence: how to formalize the common sense knowledge that most things remain the same from one moment to the next. Much attention has been focused on this problem and many technical solutions have been proposed. Yet little is understood about this problem at a fundamental level. This talk identifies three important problems that arise in temporal reasoning and discusses their solution in light of the proposed interpretation of defaults.