ILP through propositionalization and stochastic k-term DNF learning

Abstract

One promising family of search strategies to alleviate runtime and storage requirements of ILP systems is that of stochastic local search methods, which have been successfully applied to hard propositional tasks such as satisfiability. Stochastic local search algorithms for propositional satisfiability benefit from the ability to quickly test whether a truth assignment satisfies a formula. Because of that many possible solutions can be tested and scored in a short time. In contrast, testing whether a clause covers an example in ILP takes much longer, so that far fewer possible solutions can be tested in the same time. Therefore in this paper we investigate stochastic local search in ILP using a relational propositionalized problem instead of directly use the first-order clauses space of solutions. © Springer-Verlag Berlin Heidelberg 2007.