Automating the Refinement of Reinforcement Learning Specifications