Neural Network Approximation for Pessimistic Offline Reinforcement Learning