Q-SAT: Value Factorization with Self-Attention for Deep Multi-Agent Reinforcement Learning.