Unifying Version Space Representations: Part II

In this paper we continue our research on unification of version-space representations. We consider the open question how to unify one-sided versionspace representations. To answer the question we introduce a new family of version-space representations that can be adapted during the learning process. The family consists of two dual representations: adaptable one-sided maximal boundary sets (AOMBS) and adaptable one-sided minimal boundary sets (AOmBS). Without loss of generality the article covers in detail only the first representation. AOMBS are defined by the positive training set and the set of maximal boundary sets indexed by the elements of a particular covering of the negative training set. We show that a version space has a space of AOMBS representations each having its own applicability requirements. This space includes the basic one-sided version-space representations: the one-sided maximal boundary sets (Hirsh, 1992) and the one-sided instance-based maximal boundary sets (Smirnov et al., 2002). So, we may conclude that AOMBS unify these representations. The task of learning AOMBS is viewed as a task of identifying a proper representation within the space of AOMBS representations depending on the applicability requirements given. This is demonstrated in an example where AOMBS are used to overcome the complexity problem of the one-sided maximal