Reinforcement Learning with Adaptive Regularization for Safe Control of Critical Systems