Learning bayesian networks consistent with the optimal branching