State Abstractions for Lifelong Reinforcement Learning