admin%cogsci.Berkeley.EDU@UCBVAX.BERKELEY.EDU (Cognitive Science Program) (11/18/86)
BERKELEY COGNITIVE SCIENCE PROGRAM Cognitive Science Seminar - IDS 237A Tuesday, November 25, 11:00 - 12:30 2515 Tolman Hall Discussion: 12:30 - 1:30 2515 Tolman Hall ``Analogical and Deductive Reasoning" Stuart Russell Computer Science UC Berkeley The first problem I will discuss is that of analogical reason- ing, the inference of further similarities from known similari- ties. Analogy has been widely advertised as a method for apply- ing past experience in new situations, but the traditional approach based on similarity metrics has proved difficult to operationalize. The reason for this seems to be that it neglects the importance of relevance between known and inferred similarities. The need for a logical semantics for relevance motivates the definition of determinations, first-order expres- sions capturing the idea of relevance between generalized pro- perties. Determinations are shown to justify analogical infer- ences and single-instance generalizations, and to express an apparently common form of knowledge hitherto neglected in knowledge-based systems. Essentially, the ability to acquire and use determinations increases the set of inferences a system can make from given data. When specific determinations are unavailable, a simple statistical argument can relate similar- ity to the probability that an analogical solution is correct, in a manner closely connected to Shepard's stimulus generaliza- tion results. The second problem, suggested by and subsuming the first, is to identify the ways in which existing knowledge can be used to help a system to learn from experience. I describe a simple method for enumerating the types of knowledge (of which determinations are but one) that contribute to learn- ing, so that the monolithic notion of confirmation can be teased apart. The results find strong echoes in Goodman's work on induction. The application of a logical, knowledge-based approach to the problems of analogy and induction indicates the need for a system to be able to detect as many forms of regu- larity as possible in order to maximize its inferential capa- bility. The possibility that important aspects of common sense are captured by complex, abstract regularities suggests further empirical research to identify this knowledge.