An EDA based on Bayesian networks constructed with Archimedean copulas

Abstract

In this paper, an estimation of distribution algorithm that adopts a copula Bayesian network as probabilistic graphic model is presented. Multivariate Archimedean copula functions with one parameter are used to model the dependences between variables and the beta distribution is used to describe the univariate marginals. The learning process of the Bayesian network is assisted through a simple technique that relies on the associative property of Archimedean copulas, the use of Kendall's tau coefficient for measuring relations between variables and the relation between tau coefficients and bivariate Archimedean copulas. This paper presents the proposal, together with some initial experiments, which show encouraging results.