Glue: Enhancing Compatibility and Flexibility of Reinforcement Learning Platforms by Decoupling Algorithms and Environments.