Explanation-Based Learning in Logic Programming Extended Abstract

It has been argued in the literature that logic programming provides a uniform, expressive, and semantically clean framework for all aspects explanation-based general- ization. Previous treatments, however, are inadequate in that they do not work well in dicult problem domains such as theorem proving or formal program development, pri- marily because meta-programs for such tasks in traditional logic programming languages such as Prolog are not declarative enough. In (4) we develop a higher-order approach to explanation-based generalization in t uProlog (an extension of Prolog by the modal t u operator) and demonstrate how previously intractable generalization problems be- came feasible. In this paper we review our approach and then address the problem of assimilation of generalizations. Assimilation bridges the gap between explanation- based generalization and explanation-based learning and, we believe, is too dicult for a general solution in terms of the underlying architecture, but rather must be under the programmer's control. Our solution is to add a very limited amount of forward reasoning by extending t uProlog with two new constructs, rule and rule_ebg which can be given a clean declarative semantics (unlike assert). While these constructs are proposed and applied in the framework of Prolog and explanation-based learning, the underlying idea is general and might also be useful for declarative search control in languages like Prolog.