A Perceptual Criterion for Visually Controlling Learning

Abstract

An advantage of diagrammatic reasoning is that it avoids traversing irrelevant paths of inference by controlling search using visual scanning on diagrams. An effective procedure for this in the domain of geometry problem solving is to visually recognize relevant perceptual-chunks in the diagrams of problems, and to use such chunks to guide problem solving. In spite of this beneficial role of perceptual-chunks in diagrammatic reasoning, however, past research has not addressed the issues of acquiring a useful set of perceptual-chunks specific to the target domain, and demonstrating the utility of acquired chunks. This paper addresses these issues by devising a mechanism for learning perceptual-chunks from problem solving episodes. Its basic concept is that the learner acquires, from the problem diagram perceptual chunks each of which is an assembly of diagram elements that can be visually recognized and grouped together. Recognition rules implement this chunking criterion in the learning system PCLEARN. We show the feasibility of the criterion by presenting experimental data on the operationality and cost-effective utility of the learned perceptual chunks in the geometry domain.