“Reinforcement learning is capable of learning the correct boundary for when a model should express uncertainty or say "I don't know".”