Globally induced model trees: An evolutionary approach

Abstract

In the paper we propose a new evolutionary algorithm for induction of univariate regression trees that associate leaves with simple linear regression models. In contrast to typical top-down approaches it globally searches for the best tree structure, tests in internal nodes and models in leaves. The population of initial trees is created with diverse top-down methods on randomly chosen subsamples of the training data. Specialized genetic operators allow the algorithm to efficiently evolve regression trees. Akaike's information criterion (AIC) as the fitness function helps to mitigate the overfitting problem. The preliminary experimental validation is promising as the resulting trees can be significantly less complex with at least comparable performance to the classical top-down counterparts. © 2010 Springer-Verlag.