Pruning GP-Based Classifier Ensembles by Bayesian Networks