Constructive induction and genetic algorithms for learning concepts with complex interaction

Abstract

Constructive Induction is the process of transforming the original representation of hard concepts with complex interaction into a representation that highlights regularities. Most Constructive Induction methods apply a greedy strategy to find interacting attributes and then construct functions over them. This approach fails when complex interaction exists among attributes and the search space has high variation. In this paper, we illustrate the importance of applying Genetic Algorithms as a global search strategy for these methods and present MFE2/GA1, while comparing it with other GA-based Constructive Induction methods. We empirically analyze our Genetic Algorithm's operators and compare MFE2/GA with greedy-based methods. We also performed experiments to evaluate the presented method when concept has attributes participating in more than one complex interaction. In experiments that are conducted, MFE2/GA successfully finds interacting attributes and constructs functions to represent interactions. Results show the advantage of using Genetic Algorithms for Constructive Induction when compared with greedy-based methods.