Learning Bayesian Networks Using Evolutionary Algorithm and a Variant of MDL Score

Abstract

Deterministic search algorithm such as greedy search is apt to get into local maxima, and learning Bayesian networks (BNs) by stochastic search strategy attracts the attention of many researchers. In this paper we propose a BN learning approach, E-MDL, based on stochastic search, which evolves BN structures with an evolutionary algorithm and can not only avoid getting into local maxima, but learn BNs with hidden variables. When there exists incomplete data, E-MDL estimates the probability distributions over the local structures in BNs from incomplete data, then evaluates BN structures by a variant of MDL score. The experimental results on Alarm, Asia and an examplar network verify the validation of E-MDL algorithm.