Events

Reinforcement learning agents learn by trial and error, which is a problem when some errors are unacceptable. Shielding is a family of methods that sits between a learning agent and its environment and blocks actions that would violate a formally specified safety property, during both training and deployment. Since its introduction in 2018, shielding has grown from a grid-world technique with absolute guarantees into a broader toolkit covering probabilistic risk, partial observability, multi-agent systems, continuous control, and offline or model-free settings. This talk introduces the core idea and key papers of shielded reinforcement learning. No matter your background, you are welcome to join this 30-minute presentation, followed by 30 minutes of Q&A and open discussion.