craftax

An empirical study of reinforcement learning scaling laws for an actor-critic algorithm in the Starpilot and Craftax environments, showing that Elo-based performance metrics are more predictable than mean episode return and enabling state-of-the-art results in Craftax through compute scaling.

This paper compares four methods for training ethical reinforcement learning agents, evaluating them on per-episode violation distributions rather than mean rates, and demonstrates that per-episode guarantees can be achieved without compromising mean performance.