Dynamic Learning Bias Selection

Christopher J. Merz
ICS Dept
University of California, Irvine
Irvine, CA 92717
cmerz@ics.uci.edu
(714)725-3491

1 Introduction
Determining the conditions for which a given learning algorithm is appropriate is an open problem in machine learning. Methods for selecting a learning algorithm for a given domain (e.g. Aha, 1992; Breiman, et al, 1984) or for a portion of the domain (Brodley, 1993) have met with limited success. This paper proposes a new approach which dynamically selects a learning algorithm for each test example by observing the prediction patterns of a suite of algorithms given in a Òcross-validation history.Ó This dynamic selection of a learning algorithm, DS, frequently outperforms a cross-validation algorithm for selecting a learning algorithm and occasionally outperforms the algorithm with the best test accuracy.