Uncertainty Estimation based Intrinsic Reward For Efficient Reinforcement Learning