Controlling large scale multi-agent environments with model-based reinforcement learning