Connections between POMDPs and partially observed n-player mean-field games
Informal聽Systems聽Seminar聽(ISS) Centre聽for聽Intelligent聽Machines聽(CIM)聽and聽Groupe聽d'Etudes聽et聽de Recherche聽en聽Analyse聽des聽Decisions聽(GERAD)
*Note that this is a hybrid event*
惭别别迟颈苍驳听滨顿:听845听1388听1004听听听听听听听
笔补蝉蝉肠辞诲别:听痴滨厂厂
Speaker: Bora Yongacoglu
Abstract: In this talk, we will study a discrete-time model of mean-field games with finitely many players and partial observability of the global state, and we will describe the deep connection between such n-player mean-field games and partially observed Markov decision problems (POMDPs). We focus primarily on settings with mean-field observability, where each player privately observes its own local state as well as the complete mean-field distribution. We prove that if one's counterparts use symmetric 聽stationary memoryless policies, then a given agent faces a fully observed, time homogenous MDP. We leverage this to prove the existence of a memoryless, stationary perfect equilibrium in the n-player game with mean-field observability. We also show, by example, that the symmetry condition cannot be relaxed without loss of generality. Under narrower observation channels, in which the mean-field information is compressed before being observed by each agent, we show that the agent faces a POMDP rather than an MDP, even when its counterparts use symmetric policies.
Bio : Bora Yongacoglu is a post-doctoral fellow in the Department of Electrical and Computer Engineering at University of Toronto, where he studies learning in multi-agent systems. He received his PhD and MSc. degrees in mathematics from Queen's University, and his B.A. in mathematics and economics from 苹果淫院.