A genetic algorithm for propositionalization

Abstract

Nowadays, propositionalization is an important method that aims at reducing the complexity of Inductive Logic Programming, by transforming a learning problem expressed in a first order formalism into an attribute-value representation. This implies a two steps process, namely finding an interesting pattern and then learning relevant constraints for this pattern. This paper describes a novel genetic approach for handling the second task. The main idea of our approach is to consider the set of variables appearing in the pattern, and to learn a partition of this set. Numeric constraints are directly put on the equivalence classes involved by the partition rather than on variables. We have proposed an encoding for representing a partition by an individual, and general set-based operators to alter one partition or to mix two ones. For propositionalization, operators are extended to change not only the partition but also the associated numeric constraints.