MA-MIX: Value Function Decomposition for Cooperative Multiagent Reinforcement Learning Based on Multi-Head Attention Mechanism.