Advantages and Limitations of using Successor Features for Transfer in Reinforcement Learning