Using simulated annealing in ILP
Résumé
In Inductive Logic Programming (ILP), algorithms which are purely of the bottom-up or the top-down type encounter several problems in practice. Since these algorithms are greedy ones, they find rules in local optima depending on the rule "quality" estimation used by the algorithm. Moreover, the existence of a phase transition between over-generalized and over-specialized rules, which has recently be pointed out, explains why extracted rules may be unsatisfactory from an induction point of view. In this paper, we describe a heuristical method of induction based on simulated annealing and allowing to escape from local optima. Using a refinement operator, we get a notion of neighborhood, which together with a rule evaluation measure allow us to define a distance between clauses. We discuss the needed conditions on the refinement operator and the rule evaluation measure for making the algorithm, efficient. An application is described and experimentation results that show the benefit of the method, are presented.