Evolutionary optimization using graphical models

We have previously shown that a genetic algorithm can be approximated by an evolutionary algorithm using the product of univariate marginal distributions of selected points as search distribution. This algorithm (UMDA) successfully optimizes difficult multi-modal optimization problems. For correlated fitness landscapes more complex factorizations of the search distribution have to be used. These factorizations are used by the Factorized Distribution Algorithm FDA. In this paper we extend FDA to an algorithm which computes a factorization from the data. The factorization can be represented by a Bayesian network. The Bayesian network is used to generate the search points.