A Survey of Machine Learning Systems Integrating Explanation-Based and Similarity-Based Methods

Danyluk, Andrea Pohoreckyj

Two disparate machine learning approaches have received considerable attention. These are explanation-based and similarity-based learning. The basic goal of an explanation-based learning system is to more efficiently recognize concepts that it is already capable of recognizing. The learning process involves a knowledge-intensive analysis of an environment-provided example of a concept in order to extract its characteristic features. The basic goal of a similarity-based system, on the other hand, is to acquire descriptions that allow the system to recognize concepts it does not yet know. Although they have been applied with some success to problems in a variety of domains, both methods have clear deficiencies. Explanation-based learning assumes that a system will be provided with an explicit domain theory that is complete, correct, and tractable. This assumption is unrealistic for many complex, real-world domains. Similarity-based learning suffers because of its lack of an explicit theory. Since the two methods are complementary in nature, an obvious solution is to augment systems using one approach with techniques from the other. This survey discusses machine learning systems that integrate explanation-based and similarity-based learning methods such that one is incorporated primarily to handle a deficiency of the other. Although sufficient background material is provided that the reader need not be familiar with machine learning, general knowledge of AI is assumed.



More About This Work

Academic Units
Computer Science
Department of Computer Science, Columbia University
Columbia University Computer Science Technical Reports, CUCS-467-89
Published Here
January 11, 2012