Learning Logic Programs with Negation as Failure

Abstract

Normal logic programs are usually shorter and easier to write and understand than definite logic programs.As a consequence, it is worth investigating their learnability, if Inductive Logic Programming is to be proposed as an alternative tool for software development and Software Engineering at large.In this paper we present an extension of the ILP system TRACY, called TRACY not , able to learn normal logic programs.The method is proved to be sound, in the sense that it outputs a program which is complete and consistent w.r.t. the examples, and complete, in the sense that it does find a solution when it exists.Compared to learning systems based on extensionality, TRACY and TRACY not are less dependent on the kind and number of training examples, which is due to the intensional evaluation of the hypotheses and, for TRACY not , to the possibility to have restricted hypothesis spaces through the use of negation.