Analogical representation in temporal, spatial, and mnemonic reasoning

Abstract

The traditional Euclidean approach to problem solving in AI has always designed representations for a domain and then spent considerable effort on the methods of efficiently searching the representation in order to extract the desired information. We feel that the emphasis in problem solving should be on the automated construction of the knowledge representation and not on the searching of the representation. This thesis proposes and implements an alternative approach: that of analogical representation. Analogical representation differs from the Euclidean methodology in that it creates a representation for the data from which the acquisition of information is done by simple 'observation.' It is not our goal to propose a system that reduces the NP-hard problem of temporal reasoning to a lower complexity. Our approach simply minimizes the number of times that we must pay the exponential expense. Furthermore, the representation can encode uncertainty and unknownness in an efficient manner. This allows for 'intelligent' creation of a representation and removes the 'mindless' mechanical search techniques from information retrieval, placing the computational effort where it should be: on representation construction.