Phase transition and New Fitness Function based Genetic Inductive Logic Programming algorithm

Abstract

A new genetic inductive logic programming (GILP for short) algorithm named PT-NFF-GILP (Phase Transition and New Fitness Function based Genetic Inductive Logic Programming) is proposed in this paper. Based on phase transition of the covering test, PT-NFF-GILP randomly generates initial population in phase transition region instead of the whole space of candidate clauses. Moreover, a new fitness function, which not only considers the number of examples covered by rules, but also considers the ratio of the examples covered by rules to the training examples, is defined in PT-NFF-GILP. The new fitness function measures the quality of firstorder rules more precisely, and enhances the search performance of algorithm. Experiments on ten learning problems show that: 1) the new method of generating initial population can effectively reduce iteration number and enhance predictive accuracy of GILP algorithm; 2) the new fitness function measures the quality of first-order rules more precisely and avoids generating over-specific hypothesis; 3) The performance of PT-NFF-GILP is better than other algorithms compared with it, such as G-NET, KFOIL and NFOIL.