Minimizing Communication while Maximizing Performance in Multi-Agent Reinforcement Learning