Large-Scale Empirical Study of 160,000 Offline Policy Training Runs

An investigation of 160,000 training runs across 114 datasets reveals that hyperparameter tuning and reporting choices frequently outweigh algorithmic superiority in offline reinforcement learning.