Representation Learning For Efficient Deep Multi-Agent Reinforcement Learning