Synthesizing Safe Policies under Probabilistic Constraints with Reinforcement Learning and Bayesian Model Checking

We propose to leverage epistemic uncertainty about constraint satisfaction of\na reinforcement learner in safety critical domains. We introduce a framework\nfor specification of requirements for reinforcement learners in constrained\nsettings, including confidence about results. We show that an agent's\nconfidence in constraint satisfaction provides a useful signal for balancing\noptimization and safety in the learning process.\n