Improving Monte Carlo Tree Search Policies in StarCraft via Probabilistic Models Learned from Replay Data