Improving Probabilistic Bisimulation for MDPs Using Machine Learning