Knowledge representation issues in control knowledge learning

Abstract

Knowledge representation is a key issue for any machine learning task. There have already been many comparative studies about knowledge representation with respect to machine learning in classication tasks. However, apart from some work done on reinforcement learning techniques in relation to state representation, very few studies have concentrated on the eect of knowledge representation for machine learning applied to problem solving, and more specically, to planning. In this paper, we present an experimental comparative study of the eect of changing the input representation of planning domain knowledge on control knowledge learning. We show results in two classical domains using three dierent machine learning systems, that have previously shown their eectiveness on learning planning control knowledge: a pure ebl mechanism, a combination of ebl and induction (hamlet), and a Genetic Programming based system (evock).