First Order Learning, Zeroth Order Data

R.M.Cameron-Jones and J.R.Quinlan

Basser Department of Computer Science

The University of Sydney, NSW 2006

Australia

[email protected] and [email protected]

Abstract

The augmentation of FOIL to include the handling of data with continuous variables and data with missing values enables it to be applied to a wider range of problems, including conventional attribute-value classification problems. This paper reports the results of experiments applying this variant of FOIL to such problems in two ways: restricting theories to the conventional form in which each test involves a single attribute, and allowing an extended range of tests in which one attribute is compared to another. The experiments indicate that both kinds of learning are computationally practical by comparison with C4.5rules and that there is little difference in accuracy on the domains tried. It thus appears that FOIL has a broad range of potential applicability, but no great advantage on zeroth order learning tasks.
Topic: Machine Learning
Keywords: first order learning, continuous variables, attribute-value representation, empirical evaluation.