Learning to Share in Multi-Agent Reinforcement Learning