Although RL is highly general and scalable, the difficulty of verifying policy behaviours poses challenges for safety-critical applications. To remedy this, we propose to apply verification methods used in control theory to learned value functions. By analyzing a simple task structure for safety preservation, we derive original theorems linking value functions to control barrier functions. Inspired by this, we propose novel metrics for verification of value functions in safe control tasks, and practical implementation details that improve learning. Besides proposing a novel method for certificate learning, our work unlocks a wealth of verification methods in control theory for RL policies, and represents a first step towards a framework for general, scalable, and verifiable design of control systems.
翻译:暂无翻译