A method for finding consistent hypotheses using abstraction

Abstract

We present in this paper a method for finding target hypotheses in Inductive Logic Programming(ILP). In order to find them efficiently, we propose to use abstraction. Given an ILP problem and a hypothesis space H, we first consider an abstraction of H. An abstract space corresponds to a small subspace of H. Then we try to find hypotheses satisfying a certain condition by searching in several such abstract spaces. Since each abstract space is small, the task is not difficult. From these hypotheses, we can easily identify a hypothesis space in which all consistent hypotheses can be found. Since the obtained space is a part of the original H, we can expect that the targets are efficiently found by searching only in the part.