Doubly Robust Off-policy Value Evaluation for Reinforcement Learning