Under CopyrightSchmoeller da Roza, FelippeFelippeSchmoeller da RozaRoscher, KarstenKarstenRoscherGünneman, StephanStephanGünneman2023-10-122024-02-202023-10-122023https://doi.org/10.24406/publica-1983https://publica.fraunhofer.de/handle/publica/451614https://doi.org/10.24406/publica-198310.24406/publica-1983Improving safety in model-free Reinforcement Learning is necessary if we expect to deploy such systems in safety-critical scenarios. However, most of the existing constrained Reinforcement Learning methods have no formal guarantees for their constraint satisfaction properties. In this paper, we show the theoretical formulation for a safety layer that encapsulates model epistemic uncertainty over a distribution of constraint model approximations and can provide probabilistic guarantees of constraint satisfaction.enreinforcement learningRLartificial intelligenceAIsafetysafe AIconstrained Markov decision processCMDPpaper