Asynchronous Factorization for Multi-Agent Reinforcement Learning

Enrico Marchesini, Yuchen Xiao, Priya L. Donti, Christopher Amato

Keywords: Macro-actions, Multi-Agent Reinforcement Learning, Asynchronous Factorization.

TL;DR: We propose value factorization for asynchronous multi-agent systems, scaling up multi-agent reinforcement learning to consider asynchronous real-world agents.

Abstract: Value factorization is widely used to design high-quality, scalable multi-agent reinforcement learning algorithms. However, current methods typically assume agents execute synchronous, 1-step *primitive actions*, failing to capture the typical nature of multi-agent systems. In reality, agents are asynchronous and execute *macro-actions*---extended actions of variable and unknown duration---making decisions at different times. This paper proposes value factorization for asynchronous agents. First, we formalize the requirements for consistency between centralized and decentralized macro-action selection, proving they generalize the primitive case. We then propose update schemes to enable factorization architectures to support macro-actions. We evaluate these asynchronous factorization algorithms on standard macro-action benchmarks, showing they scale and perform well on complex coordination tasks where their synchronous counterparts fail.

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: 11608