Reward Generation for Model-Free Reinforcement Learning from Formal Specifications.