Genetic approach to constructive induction based on non-algebraic feature representation

Abstract

The aim of constructive induction (CI) is to transform the original data representation of hard concepts with complex interaction into one that outlines the relation among attributes. CI methods based on greedy search suffer from the local optima problem because of high variation in the search space of hard learning problems. To reduce the local optima problem, we propose a CI method based on genetic (evolutionary) algorithms. The method comprises two integrated genetic algorithms to construct functions over subsets of attributes in order to highlight regularities for the learner. Using non-algebraic representation for constructed functions assigns an equal degree of complexity to functions. This reduces the difficulty of constructing complex features. Experiments show that our method is comparable with and in some cases superior to existing CI methods.