Abstract

Classifier systems are knowledge-based learning algorithms that take training instances as input and produce a set of rules as output. The classifier systems focused on in this paper represent the knowledge they learn in the form of decision trees, and are built upon Quinlan's [ 1986] inductive algorithm ID3.

The ID3 algorithm suffers from the inability to easily and effectively handle domains with numeric-valued attributes. Numeric attributes are those whose values are taken from a continuous domain or from a domain with a large number of discrete values. A number of approaches have been developed for extending ID3 to handle numeric attributes. Additionally, a number of approaches have been suggested for optimizing the accuracy and/or generalization of trees produced by the basic ID3 system. However, research to date has not focused on combining these approaches to optimize accuracy and generalization ability in classification systems which handle numeric attributes.

This paper examines the various approaches for handling numeric attributes and for optimizing accuracy and generalization. An empirical comparison of combinations of these approaches is presented in order to evaluate the accuracy and generalization ability of the various numeric classifier systems.

Department(s)

Computer Science

Second Department

Mathematics and Statistics

Comments

The first Author is a Graduate Student.

This report is substantially the M.S. thesis of the first author, completed May, 1992.

This thesis has been prepared in the format used by the Association for Computing Machinery (ACM). Pages 1-45 will be presented for publication in the journal Communications of the ACM.

Keywords and Phrases

Automated Induction, Machine Learning, Knowledge Representation

Report Number

CSc-92-13

Document Type

Technical Report

Document Version

Final Version

File Type

text

Language(s)

English

Rights

© 1992 University of Missouri - Rolla, All rights reserved

Publication Date

1 May, 1992

Download

Share

 
COinS