Crowd Simulation Via Multi-Agent Reinforcement Learning