Safe Reinforcement Learning with Model Uncertainty Estimates