Chongjie Zhang

Assistant Professor

Institute for Disciplinary Information Sciences

Tsinghua University 

Cooperation is key to the success of humankind. As the scale of complexity increases in AI applications, we must endow AI agents with similar cooperative abilities to achieve beyond what they can do alone. In this talk, I will discuss how AI agents can autonomously learn to cooperate and coordinate with other agents. I will first introduce a cooperative multi-agent reinforcement learning (MARL) framework with factorization structures and provide formal analyses to reveal its theoretical properties. Based on these theoretical insights, I will present novel factored MARL methods that achieve state-of-the-art performance in challenging benchmarks. I will then briefly discuss approaches that are built upon factored MARL for addressing practical challenges of cooperative MARL, including learning efficiency, partial observability, and coordinated exploration. I will finish the talk by outlining some promising directions of future work.