Systematic Predicate Invention in Inductive Logic Programming

We propose in this paper a new approach for learning predicate deenitions from examples and an initial theory. The particularity of this approach consists in inventing a new predicate at most steps of learning; once the learning task is ended, most invented predicates are removed by unfolding techniques. Nevertheless, some predicates remain in the learned deenitions, either because they enable to simplify the program, or because their deenitions are recursive and the program could not have been learned without inventing these predicates. Moreover, when a new predicate symbol is introduced , a speciication for this predicate is built; this speciication is both incomplete and imprecise, what we modelize by introducing the notion of-interpretation. It is worth noted that the proposed method can be used, even when the target concepts are also incompletely deened by-interpretations. The method has been implemented in the system, called SPILP, which has been successfully tested for inventing predicates which simplify the learned programs as well as for inventing recursively deened predicates.