Addressing Extrapolation Error in Deep Offline Reinforcement Learning