Sample Complexity Reduction via Policy Difference Estimation in Tabular Reinforcement Learning