Integrating similarity-based and explanation-based learning

Abstract

Recently, there have been various attempts to combine the strengths of similarity-based learning (SBL) and explanation-based learning (EBL) in a single learning system. We describe a graph-based learning method called Graph Induction, which is based on the graphical representation of a formal lattice and supports both supervised and unsupervised learning. The method integrates SBL with a weak form of EBL in such a way that the two mechanisms become totally blended. The result is a unified algorithm with both SBL and EBL involved in each step. The domain theory is generated and/or extended as SBL proceeds and employed immediately, through EBL, to guard further learning and thus control the size of the lattice which otherwise has the potential for increasing exponentially.