AIRS in the AIR

• 

  查宏远

  香港中文大学（深圳）校长学勤讲座教授、数据科学学院执行院长、AIRS 机器学习与应用中心主任

  执行主席

  嘉宾介绍

  个人简介
• 

  王趵翔

  香港中文大学（深圳）数据科学学院助理教授、AIRS 机器学习与应用中心副研究员

  主持人

  嘉宾介绍

  个人简介
• 

  张崇洁

  清华大学交叉信息研究院助理教授

  Cooperative Multi-Agent Reinforcement Learning with Factored Value Functions

  嘉宾介绍

  张崇洁，清华大学交叉信息研究院助理教授，博士生导师。2011年在美国麻省大学阿默斯特分校获计算机科学博士学位，后在麻省理工学院从事博士后研究。目前的研究兴趣主要在人工智能、强化学习、多智能体系统等领域。

  个人简介

  Collaboration is indispensable for solving complex tasks. Learning to collaborate effectively is one of the key problems in artificial intelligence. Cooperative multi-agent reinforcement learning (MARL) potentially provides a promising solution, but faces two fundamental challenges: scalability and credit assignment. In this talk, I will discuss a MARL paradigm with factored value functions to address these challenges. I will first present formal analysis on factored value learning, revealing its implicit credit assignment mechanism and properties of convergence and optimality. Inspired by these theoretical insights, two novel MARL methods will then be introduced with linear and non-linear value factorization, respectively, which achieves state-of-the-art performance. Building on factored MARL, I will also briefly discuss approaches for addressing other challenges of cooperative MARL, such as learning efficiency, partial observability, and exploration.

• 

  卢宗青

  北京大学计算机学院助理教授

  Fully Decentralized Multi-Agent Reinforcement Learning

  嘉宾介绍

  Zongqing Lu is currently a BOYA assistant professor in School of Computer Science at Peking University. He is also affiliated with Institute of AI at Peking University and Beijing Academy of Artificial Intelligence. His current research focuses on reinforcement learning and AI systems.

  个人简介

  The main research of multi-agent reinforcement learning (MARL) focuses on the paradigm of centralized training with decentralized execution (CTDE). Another potential paradigm is fully decentralized learning, which is less investigated but better for robustness, scalability, and generalibility. However, current fully decentralized learning methods do not even have convergence guarantee. In this talk, I will present our recent studies on fully decentralized learning algorithms with convergence guarantee, including value-based, actor-critic, and model-based methods.

视频回顾