Wenhao Li
Post Doctoral Fellow
Institute of Robotics and Intelligent Manufacturing,
the Chinese University of Hong Kong, Shenzhen.
Former Researcher at Tencent AI Lab.
Ph.D. from East China Normal University,
advised by Prof. Aimin Zhou and Prof. Hongyuan Zha,
co-advised by A.P. Bo Jin and A.P. Xiangfeng Wang.
Post Doctoral Fellow
Institute of Robotics and Intelligent Manufacturing,
the Chinese University of Hong Kong, Shenzhen.
Former Researcher at Tencent AI Lab.
Ph.D. from East China Normal University,
advised by Prof. Aimin Zhou and Prof. Hongyuan Zha,
co-advised by A.P. Bo Jin and A.P. Xiangfeng Wang.
My primary research focus encompasses the resolution of intricate multi-agent collaboration dilemmas within diverse real-world decision-making tasks using multi-agent reinforcement learning (MARL) algorithms. My published works chiefly comprise two facets: methodologies and practical applications:
- Concerning methodologies, I strive to devise MARL algorithms exuding superior robustness, scalability, and transferability. Additionally, I endeavor to incorporate generative models (GFlowNet, diffusion models, etc.) within MARL, aiming to augment the expressive capacity of neural network policies and bolster algorithmic competence in tackling high-dimensional problems.
- In the realm of applications, I am dedicated to remodeling a variety of real-world issues from the perspective of stochastic games, encompassing image segmentation, multi-agent path-finding, and precision agriculture, utilizing MARL algorithms for resolutions. Concurrently, my interests also lie in the fusion of MARL with computational social science, formulating and solving quintessential social dilemmas in social sciences through a stochastic game lens.
- LLM + MARL (AI Agents): Through the integration of the vast human knowledge encapsulated within LLMs into the MARL paradigm, the colossal exploration space hitherto demanded by end-to-end learning can be significantly mitigated, thereby substantially enhancing the sample efficiency of algorithms.
- In-context MARL: Evidence suggests that the formidable zero-shot generalization capability of LLMs partly stems from their in-context learning capabilities. By endowing the reinforcement learning paradigm with in-context learning faculties, I seek to bolster decision-making proficiencies across various specialized, cooperative tasks in diverse domains.