Know What It Does Not Know: Improving Offline Deep Reinforcement Learning With Uncertainty Estimation