International Joint Conference On Theoretical Computer Science – Frontier of Algorithmic Wisdom

August 15-19, 2022, City University of Hong Kong, Hong Kong

 

Invited Speakers

Track C


Decision Structure in Decentralized Multi-Agent Learning

Yali Du

King’s College London

Abstract:
Many real-world problems, such as traffic light control, connected autonomous vehicles, and multiplayer games, can be naturally formulated into multi-agent learning problems. While centralized controllers are an off-the-shelf choice enabling the use of various existing reinforcement learning algorithms, a long-standing challenge is the scalability of the algorithms with exponential growing state-actions spaces with the number of agents. Especially in large scale multi-agent systems, centralized algorithms are hard to scale, and fully decentralized algorithms are favored. While in reality some decision makers might not be influenced by distant agents, such as in traffic lights, we are inspired to investigate the structure of multi-agents systems and understand how it will promote the design of better decentralized learning algorithms.
In this talk, we will discuss factorizability of multi-agent systems including the transition dynamics and action coordinations, and propose novel fully decentralized multi-agent learning algorithms. Theoretically we find that the localized policy gradient is a close approximation to true policy gradients and provide the conditions for monotonic improvement under factorized model-based policy optimization. Empirically, we evaluate our algorithm on intelligent transportation systems, demonstrating the high sample efficiency and superior performance.

Bio:
Yali Du is a Lecturer at King’s College London and a member of the Distributed Artificial Intelligence Group. Prior to joining King’s, she was a postdoctoral research fellow at University College London. She obtained her Ph.D. from University of Technology Sydney in 2019. Her research interest lies in machine learning and reinforcement learning, especially in the topics of multi-agent learning, policy evaluation, and applications to Game AI, data science and wide decision-making tasks. Her research output has been widely published in prestigious venues including ICML, NeurIPS, ICLR, IJCAI, AAMAS, WWW, etc.
Personal website: https://yalidu.github.io/