Development of a multi-strategy theory revision

Abstract

This thesis proposes and implements a new Theory Revision System. The Theory Revision problem is defined as the problem of how best to go about revising a knowledge base on the basis of a collection of examples, some of which expose inaccuracies in the original knowledge base. This problem entreats a thorough investigation of the following machine learning field of study: Combining Inductive and Analytical learning. The problem of theory revision has been studied for quite some time and various systems have been proposed. On one hand, there are successful theory revision systems like EITHER and PTR, which combirie Inductive and Analytical Learning. On the other hand, there are mention-worthy systems like KBANN, TANGENTPROaPn,d EBNN that use imperfect domain theories together with given training set of data. The new system is built by incorporating Version Spacebased Incremental Probabilistic Evidence Combination method and Integrated AnalyticallEmpirical method. The proposed system is constructed to maXImIze preservation of already gained meaningful information. To our knowledge, Version Space-based approach has not been applied for theory revision problem as yet. Experimental results show that the performance of the new system is comparable with other fairly successful systems.