Provable Learning for DEC-POMDPs: Factored Models and Memoryless Agents

Jingchu Gai, Qinghua Liu, Zhuoran Yang, Chi Jin

Keywords: decentralized POMDP, provably efficient algorithm

TL;DR: We present sample-efficient algorithms for DEC-POMDP from a theoretical perspective.

Abstract: This paper studies cooperative Multi-Agent Reinforcement Learning (MARL) under the mathematical model of Decentralized Partially Observable Markov Decision Process (DEC-POMDP). Despite the empirical success of cooperative MARL, its theoretical foundation, particularly in the realm of provable learning of DEC-POMDPs, remains limited. In this paper, we first present a hardness result in theory demonstrating that, without additional structural assumptions, learning DEC-POMDPs requires several samples that grows exponentially with the number of agents in the worst case, which is also known as the curse of multiagency. This motivates us to explore important subclasses of DEC-POMDPs for which efficient solutions can be found. Specifically, we propose new algorithms and establish sample-efficiency guarantees that break the curse of multiagency, for finding both local and global optima in two important scenarios: (1) when agents employ memoryless policies, selecting actions based solely on their current observations; and (2) when a factored structure is present, which enables key properties similar to value decomposition in VDN or Qmix.

Primary Area: reinforcement learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 8796