SteinGate: Tail-Sensitive Safe Reinforcement Learning via Stein Discrepancy

arXiv cs.AI · 18 hours ago

Researchers introduced SteinGate, a method for safe reinforcement learning that detects rare catastrophic events by using Kernelized Stein Discrepancy instead of traditional expected cost bounds. The approach uses a non-parametric safety certificate that compares observed policy costs against a safe reference distribution, switching to recovery behavior when costs deviate. SteinGate reduced constraint violations during training while maintaining competitive performance on continuous-control benchmarks compared to existing methods.