Learning Meta Representations for Agents in Multi-Agent Reinforcement Learning