Research work · 1994

Learning Perceptually Chunked Macro Operators

M Suwa, H Motoda

Machine Intelligence 13

Abstract

Abstract In previous studies search control knowledge has been acquired using explanation-based learning (EBL) techniques. These learn goal-oriented control knowledge by explaining how a decision at a control decision node leads to the goal. In the domain of geometry problem-solving, however, this leads to knowledge which is neither sufficiently general nor sufficiently operational. This paper addresses an alternative form of search control knowledge in which the search is controlled at each decision node in such a way that a problem solver can locally recognize relevant ‘perceptual chunks’. Previously the effectiveness of perceptual chunks as control knowledge has been reported in a geometry domain. In this paper, we propose a new chunking technique, which acquires, as a chunk, an assembly of diagram elements that can be recognized and grouped together with the control decision node. In order to implement this chunking criterion a learner, Pclearn, employs recognition rules, domain specific knowledge describing necessary conditions for a domain object to be recognizable. Experiments in a geometry domain show that the set of learned knowledge exhibits higher operationality than EBL macro-operators. They also suggest that the Pclearn chunking technique can be a powerful method for obtaining a small and highly organized set of domain-specific perceptual chunks if augmented with a mechanism for dynamically managing the utility of each chunk.

DOI: 10.1093/oso/9780198538509.003.0016