机器学习与数据科学博士生系列论坛（第五十六期）—— A Unified Stochastic Approximation Framework for Learning in Games

摘要：
A key question for online learning in games is whether the players eventually settle down to a stable profile from which no player has an incentive to deviate, i.e., whether the players' learning process converges to a Nash equilibrium. The answer is positive for specific types of games such as two-player zero-sum finite games, monotone, smooth and potential games, while for general games it would be negative. Therefore, a natural question is to characterize the sets of actions that are stable and attracting under a given learning process.

In this talk, we follow the recent work of Mertikopoulos et al. [2023] to introduce a stochastic approximation framework for analyzing the long-run behavior of learning in games. The framework incorporates a wide range of learning algorithms, including gradient-based methods, multiplicative weight algorithms for learning in finite games, optimistic and bandit variants of the above, etc. Moreover, a range of criteria for identifying classes of Nash equilibria and sets of action profiles that are attracting with high probability will also be discussed.

论坛简介：该线上论坛是由张志华教授机器学习实验室组织，每两周主办一次（除了公共假期）。论坛每次邀请一位博士生就某个前沿课题做较为系统深入的介绍，主题包括但不限于机器学习、高维统计学、运筹优化和理论计算机科学。