A one-day workshop on multi-agent decision-making and reinforcement learning, particularly in the presence of decentralized information

As exemplified by the celebrated Witsenhausen's counterexample in control theory, the hardness of multi-agent decision-making is known to increase sharply when the agents have decentralized and asymmetric information. The hardness has thus led to many other fundamental results in decentralized stochastic control, such as the categorization of Information Structure of the decentralized multi-agent systems on who knows what and when, and structural results on the optimization/computational aspects of the problem, given model knowledge. On the other hand, inspired by its tremendous empirical successes, e.g., AlphaGo, robotics, autonomous driving, and large-language-model alignment, reinforcement learning (RL) has recently regained increasing attention in both the controls and the machine learning communities. Notably, many empirical successes of RL naturally involved multiple agents, inspiring the recent rapid advances on the theoretical foundations of multi-agent RL (MARL). However, most existing advances on MARL theory concentrated on the stochastic/Markov game model, and required access to the global state of the environment. The impact of information decentralization has been much less explored, especially from the lens of Information Structure, and its impact on the statistical tractability/efficiency of MARL (together with computational tractability). The proposed workshop intends to bridge the above gaps between decentralized stochastic control, dynamic games, multi-agent/networked control systems, statistics, and RL theory, by bringing together an excellent collection of speakers toward a unified theme.

• Are you a PhD student/postdoc working on Reinforcement Learning, Statistics, and Game Theory?

• We will have a session dedicated to you to present your work as lightning talks and/or posters!

• Please fill in this Google Form if you are interested in presenting

 If you have any questions, please do not hesitate to contact us via email.