Formulating concepts and analogies according to purpose
Smadar T. Kedar-Cabelli · 1988
The dissertation investigates open problems in machine learning, a subarea of artificial intelligence (AI). In particular, we focus on issues within the explanation-based generalization (EBG) framework, a framework for producing deductive generalizations from single examples. We begin by presenting a domain independent EBG method, and an associated algorithm and implementation (PROLOG-EBG). The algorithm demonstrates the close relationship between EBG and resolution theorem-proving. Despite recent progress, EBG methods exhibit an important limitation: they are incapable of determining which target concepts are useful ones to acquire. More robust generalizers must be able to automatically determine which concepts to acquire based on the purpose of the learning, since concepts acquired for one purpose may not be appropriate for another. Our notion of the purpose of the learning is to acquire concepts which will benefit an associated performance system. Two open issues become apparent once EBG is associated with a performance system: How can EBG acquire target concepts and definitions appropriate for the performance system? Further, could the acquired target concept definitions be used to improve subsequent performance? We focus and investigate these issues in the context of a specific type of performance system--a state-space planner. Our approach is to provide EBG with explicit knowledge of the planner and specific planning task. The purposive concept formulation and purposive explanation replay methods, respectively, provide solutions to the open problems. We provide experimental support for these methods in the form of prototype systems. The results confirm that a learning system can formulate concepts and analogies sensitive to the purpose of the learning in restricted planning situations. We describe further extensions suggested by these results.