Temporal Abstraction in Reinforcement Learning with the Successor Representation