PhD Dissertation Proposal by GENG Minghong | Scaling up Cooperative Multi-agent Reinforcement Learning

ABOUT THE TALK

Cooperative multi-agent reinforcement learning (MARL) aims to develop effective collaborative behaviors among multiple agents to accomplish complex tasks. While current MARL research predominantly focuses on multi-agent problems with small populations, there remains a significant research gap in addressing real-world scenarios characterized by large agent populations and extended decision-making horizons. This dissertation proposal aims to investigate the scalability challenges of cooperative MARL in two key dimensions: structurally increasing the number of cooperative reinforcement learning agents and temporally extending the time horizon and complexity of multi-agent problem domains. The proposal begins with a comprehensive survey of the existing literature on scaling up multi-agent reinforcement learning systems in the above two directions. For tackling the scalability challenge in multi-agent systems, this proposal presents a generalizable hierarchical multi-agent framework incorporating heterogeneous learning methods including self-organizing neural networks and MARL. To assess the efficacy of existing MARL methods on large-scale problems, this proposal introduces a multi-objective MARL benchmark. This benchmark is designed to evaluate current MARL methodologies in scenarios characterized by long-horizon, multiple objectives, and sequential subtask curricula. Furthermore, to enhance the capabilities of hierarchical multi-agent systems in tackling large-scale problems including the long-horizon benchmark above, this proposal discusses using large language models (LLMs) as agents' policy models to expand agent's sensing, planning, and reasoning abilities under a hierarchical framework. By addressing these challenges, this dissertation seeks to contribute novel MARL methods for addressing complex multi-agent problems more effectively.

ABOUT THE SPEAKER

GENG Minghong is a Ph.D. candidate at SCIS, under the supervision of Prof. TAN Ah Hwee. Minghong's research focuses on multi-agent reinforcement learning. He is currently exploring hierarchical approaches for scaling up multi-agent learning systems. Minghong is a member of the Neural and Cognitive Computing Group. Please visit his homepage https://gengminghong.github.io/ for more information.