ATransformational Analysis of Expensive Chunks.pdf

A Transformational Analysis of Expensive Chunks Jihie Kim and Paul S. Rosenbloom Information Sciences Institute and Computer Science Department University of Southern California 4676 Admiralty Way Marina del Rey, CA 90292, U.S.A. jihie@isi.edu, rosenbloom@isi.edu (310) 822-1510 (x769) Fax: (310) 823-6714 Key words: machine learning, utility problem, expensive chunks, Soar, EBL Abstract Many learning systems must confront the problem of run time after learning being greater than run time before learning. This utility prob- lem has been a particular focus of research in explanation-based learning (EBL). This paper shows how the cost increase of a learned rule in an EBL system can be analyzed by characterizing the learning process as a sequence of transformations from a problem solving episode to a learned rule. The analysis of how the cost changes through the transformations can be a useful tool for revealing the sources of cost increase in the learning system. Once all of the sources are revealed, by avoiding these sources, the learned rule will never be expensive. That is, the cost of the learned rule will be bounded by the problem solving. We performed such a transformational analysis of chunking in Soar. The chunking process has been decomposed into a sequence of transformations from the problem solving to a chunk. By analyzing these transformations, we have identied a set of sources which can make the output chunk expensive. 1 1 Introduction Many learning systems must confront the problem of run time after learning being greater than run time before learning. This utility problem has been a particular focus of research in explanation-based learning (EBL). There have been approaches which are useful for producing cheaper rules [1, 2, 3, 4, 5, 6] or ltering out expensive rules [2, 7, 8, 9]. However, these approaches can- not generally guarantee that the cost of using the learned rules will always be bounded by the cost of the problem solving from which they are l