Achieving Sample and Computational Efficient Reinforcement Learning by Action Space Reduction via Grouping