Using empirical subsumption to reduce the search space in learning

Abstract

In the traditional learning framework, hypothesis that are not equivalent with respect to the standard subsomption relation can be equivalent from the learning’s point of view. We define in this paper a new subsumption relation, called empirical subsumption, that allows to take into account this fact. This new subsomption relation is then used to define a particular kind of search space reduction that do not reduce the class of learnable concepts. Then, we show that theses theoretical results can be applied when the knowledge representation formalism is the conceptual graph formalism.