A framework for incremental learning of logic programs

Abstract

In this paper, a framework for incremental learning is proposed. The predicates already learned are used as background knowledge in learning new predicates in this framework. The programs learned in this way have nice modular structure with conceptually separate components. This modularity gives the advantages of portability, reliability and efficient compilation and execution. Starting with a simple idea of Miyano et al. [21,22] for identifying classes of programs which satisfy the condition that all the terms occurring SLD-derivations starting with a query are no bigger than the terms in the initial query, we identify a reasonably big class of polynomial time learnable logic programs. These programs can be learned from a given sequence of examples and a logic program defining the already known predicates. Our class properly contains the class of innermost simple programs of [32] and the class of hereditary programs of [21,22]. Standard programs for gcd, multiplication, quick-sort, reverse and merge are a few examples of programs that can be handled by our results but not by the earlier results of [21,22, 32].