The role of default representations in incremental learning

Abstract

This paper contributes to the state-of-the-art in two major ways. First, it improves on Wrobel's work on provably minimal specializations by presenting an operator which is provably minimal even in the case of iterated sequences of specializations. The operator also benefits from the advantages of lazy evaluation and deferred choice. Second, it provides a semantic basis for inducing default theories and presents an incremental learning algorithm which is sound with respect to these semantics. The only other approach to learning default theories that we are aware of is that of Dimopoulos and Kakas [3]. Our work differs significantly from their's because our setting is incremental while their's is not. The underlying default logic in their work also differs significantly from ours. We believe that our work will provide the starting point for implemented systems that learn default theories and thence, fielded practical applications.