An analytical comparison of some rule-learning programs

Reviewer: Kathleen M. Swigger

This paper attempts to develop a theoretical framework for the field of machine learning, specifically rule-learning programs. The authors accomplish this task by analyzing and comparing several AI programs in an effort to extract an algorithmic classification of techniques for rule-learning; to identify their range; to locate their flaws; to establish the relationship between them; and to extend their range. More specifically, the authors examine the work of Brazdil (ELM), Langley (AMBER), Mitchell et al. (LEX), Quinlan (ID3), Shapiro (MIS), Waterman (P), and an extension of Winston's work made by Young, Plotkin, and Linz. To aid in this comparison, the authors discuss the various AI programs using a uniform notation. By using this formalism, the reader is able to get a very clear idea of how the various systems differ (and are the same) in their approach and implementations. The authors have identified the following rule-learning techniques: criticism techniques, conjunctive modification techniques, disjunctive modification techniques, rule-creation techniques, and description space modification. This classification differs markedly from that of Smith et al. [1], who concentrate on the architecture of learning programs, and a more descriptive survey that appears in [2]. Although the authors clearly delineate how the systems fit into these classifications, they are best in their discussion of the differences between Focusing and the Candidate Elimination Algorithm, both conjunctive learning techniques. This is, perhaps, the most insightful section in the entire paper. By explaining how the two methods treat the same problem, the authors are able to demonstrate how the methods differ with respect to rule modification. It would have been nice if this same clarity of explanation had been carried over to the “Creating New Rules” section. Understandably, such a task might have required an entire book. Artificial Intelligence is a young science that desperately needs scholars who are willing to synthesize and analyze its paradigms. This paper analyzes various algorithms used in machine learning. Thus, it makes a very vital contribution to both the specific field of Artificial Intelligence and to the broader area of Computer Science.