Agnostic System Identification for Model-Based Reinforcement Learning