Learning to school in dense configurations with multi-agent deep reinforcement learning