Variable Transformations in Estimation of Distribution Algorithms

Davide Cucci, Luigi Malagò, Matteo Matteucci

Abstract: In this paper we address model selection in Estimation of Distribution Algorithms (EDAs) based on variables trasformations. Instead of the classic approach based on the choice of a statistical model able to represent the interactions among the variables in the problem, we propose to learn a transformation of the variables before the estimation of the parameters of a fixed model in the transformed space. The choice of a proper transformation corresponds to the identification of a model for the selected sample able to implicitly capture higher-order correlations. We apply this paradigm to EDAs and present the novel Function Composition Algorithms (FCAs), based on composition of transformation functions, namely I-FCA and Chain-FCA, which make use of fixed low-dimensional models in the transformed space, yet being able to recover higher-order interactions.