Ilp-95, Leuven Mult Icn: an Empirical Multiple Predicate Learner

In this paper, we are interested in empirical multiple predicate learning. The rst solution to this problem that consists in putting together the deenitions obtained by a single predicate learning system is rarely interesting. We explain why and we show how a single predicate learning system has been extended to a multiple predicate learning system called mult icn, which learns deenite logic programs. Our system is based on the notion of exten-sional coverage but during the construction of the program, it builds a set of recursive dependencies that gives information about the mutually recur-sive calls. It has therefore two main advantages. First, it ensures that the learned program is globally consistent and complete, i.e., the learned program does not only extensionally cover the positive examples and reject the negative ones but it does prove that positive examples are true and negative ones are false in the semantics of the learned program. Secondly it uses only the knowledge given by the user to induce deenitions and it is based on an acceptability rate; both enable to reduce the innuence of the order the predicates are learned.