Robust Reinforcement Learning via Leveraging Historically Optimal Policy With Regulation of Performance

This research proposes HORP, a method to make reinforcement learning (RL, where AI systems learn by trial-and-error to maximize rewards) more resistant to adversarial attacks (manipulations designed to fool the AI). HORP improves robustness by using a previously learned optimal policy (the best strategy found so far) to guide learning, creating diverse attack scenarios, and adjusting how much uncertainty to introduce during training to help the AI defend itself better.