Greedy based Value Representation for Optimal Coordination in Multi-agent Reinforcement Learning