Shields for Safe Reinforcement Learning
Bettina Könighofer, Roderick Bloem, Nils Jansen, Sebastian Junges, Stefan Pranger · Communications of the ACM · 2025
Reinforcement learning (RL) is a prominent machine learning technique used to optimize an agent’s performance in potentially unknown environments. Despite its popularity and success, RL lacks safety guarantees, both during the learning phase and deployment. This paper reviews a runtime enforcement method called shielding that ensures provable safety for RL. We describe the underlying models, the types of guarantees that can be delivered, and the process of computing shields. Furthermore, we describe several techniques for integrating shields into RL, discuss the advantages and potential drawbacks of this integration, and highlight the current challenges in shielded learning.