Abstract Reward Processes: Leveraging State Abstraction for Consistent Off-Policy Evaluation