On Valid and Invalid Methodologies for Experimentala Evaluations of EBL

Abstract

A number of experimental evaluations of explanation-based learning (EBL) have appeared in the literature on machine learning. Closer examination of experimental methodologies used in the past reveals certain methodological flaws that call into question the conclusions drawn from these experiments. This paper illustrates some of the more common methodological problems, proposes a novel experimental framework for future empirical studies of EBL, and presents an example of an experiment performed within this new framework.