PCLEARN: A Computer Model for Learning Perceptual Chunks

Issue title: Machine Learning

Article type: Research Article

Authors: Suwa, Masaki | Motoda, Hiroshi

Affiliations: Advanced Research Laboratory, Hitachi Ltd., 2520, Hatoyama, Saitama, 350-03, Japan

Abstract: Acquiring search control knowledge of high utility is essential to reasoners in speeding up their problem-solving performance. In the domain of geometry problem-solving, the role of “perceptual chunks”, an assembly of diagram elements many problems share in common, in effectively guiding problem-solving search has been extensively studied, but the issue of learning these chunks from experiences has not been addressed so far. Although the explanation-based learning technique is a typical learner for search control knowledge, the goal-orientedness of its chunking criterion leads to produce such search control knowledge that can only be used for directly accomplishing a target-concept, which is totally different from what perceptual-chunks are for. This paper addresses the issues of acquiring domain-specific perceptual-chunks and demonstrating the utility of acquired chunks. The proposed technique is that the learner acquires, for each control decision node in the problem-solving traces, a chunk which is an assembly of diagram elements that can be visually recognizable and grouped together with the control decision node. Recognition rules implement this chunking criterion in the learning system PCLEARN. We show the feasibility of the proposed technique by investigating the cost-effective utility of the learned perceptual chunks in the geometry domain, and also discuss the potential for the technique being applied to other domains.

DOI: 10.3233/AIC-1994-7204

Journal: AI Communications, vol. 7, no. 2, pp. 114-125, 1994