Markov Potential Game

Markov Potential Games (MPGs) are a subclass of Markov games where the multi-agent interactions can be represented by a single potential function, simplifying analysis and facilitating the search for Nash equilibria (NE). Current research focuses on developing and analyzing efficient algorithms, particularly independent policy gradient methods and mirror descent variants, to find NEs, especially in large-scale settings with many agents or large state spaces. This work is significant because it addresses challenges in multi-agent reinforcement learning, offering theoretical guarantees and practical algorithms for solving complex problems like sensor network localization and network load balancing. The development of provably convergent algorithms with improved scaling properties is a key area of ongoing investigation.