Offline Reinforcement Learning for Plasma Control in Nuclear Fusion: Codebase and Benchmark
Researchers have introduced RL4F, a new benchmark designed to standardize the evaluation of offline reinforcement learning for plasma control in nuclear fusion. This benchmark utilizes historical data from the DIII-D tokamak to create realistic control tasks, addressing the challenge of costly and risky online experimentation. The study found that offline model-based RL methods generally performed best, though no single approach excelled across all tasks, emphasizing the need for effective dynamics modeling in complex fusion control scenarios. The codebase, datasets, and evaluation framework have been released to encourage further research in both fusion control and offline RL algorithm development. AI
IMPACT Standardizes evaluation for offline RL in fusion, potentially accelerating progress in both fields.