Off-campus UMass Amherst users: To download dissertations, please use the following link to log into our proxy server with your UMass Amherst user name and password.
Non-UMass Amherst users, please click the view more button below to purchase a copy of this dissertation from Proquest.
(Some titles may also be available free of charge in our Open Access Dissertation Collection, so please check there first.)
Predicting the effect of instance representations on inductive learning
Abstract
The ability to generalize from examples depends on the algorithm employed for learning and the instance representation used to describe the examples to the learning algorithm. Recently a considerable effort has been devoted to the design of algorithms that generalize well in a large number of situations. However, the effect of the instance representation on the accuracy of the generalizations made by the learning algorithm is poorly understood. This dissertation describes how the duality between finding a compact description for the examples and generalizing from the examples can be utilized to determine the suitability of an instance representation for a learning algorithm. In particular, a heuristic algorithm to compare representations, called ACR, is described. Given a learning algorithm, a set of examples, and alternative instance representations for the examples, ACR attempts to identify the instance representation that will enable the learning algorithm to produce the most accurate hypothesis. For each instance representation, ACR estimates the minimum number of bits with which the learning algorithm can express the examples. Experiments with a variety of learning tasks show that ACR is effective in ranking representations. In addition, ACR was found to rank representations faster than rankings obtained by directly estimating the accuracy of the hypotheses produced with different representations. The conclusion that one representation is better than another because it improves the compressibility of the examples suggests that different methods for improving the compressibility of the examples should result in different techniques for improving representations. Two types of representation change are identified by analyzing the reason for change in compressibility of the examples with the different representations.
Subject Area
Computer science
Recommended Citation
Saxena, Sharad, "Predicting the effect of instance representations on inductive learning" (1991). Doctoral Dissertations Available from Proquest. AAI9207455.
https://scholarworks.umass.edu/dissertations/AAI9207455