Instance-Based and Generalization-Based Learning Procedures Applied To Solving Integration Problems.

Abstract

The majority of machine learning systems produce representations that contain generalizations derived from training examples but recent research has led to the development of instance-based methods that produce representations composed of sets of examples. Instance-based methods are believed to offer advantages both in the resources required for learning and the power of the resulting representation. However there has been little experimental evidence to support the view that instance-based learning can produce better representations than generalization-based approaches. In this paper we describe a comparative study of instance-based and generalization-based learning applied to the domain of symbolic integration. This domain differs from those chosen for earlier research in instance based methods in that examples cannot be described as N-tuples of feature values and the significance of individual features is highly context sensitive. We describe two types of similarity functions which can be used for this type of example. The results obtained demonstrate that instance-based learning produces a better representation and it is shown that this is a consequence of the inability of generalizations to adequately represent the complexities of the domain. We conclude that complex highly structured domains might prove to be one of the most profitable areas for application of instance-based methods.