Addressing Extrapolation Error in Multi-Agent Reinforcement Learning

Yueheng Li, Guangming Xie

Keywords: cooperative multi-agent reinforcement learning, CTDE, value factorization, extrapolation error

Abstract:

Cooperative Multi-Agent Reinforcement Learning (MARL) has become a critical tool for addressing complex real-world problems.
However, scalability remains a significant challenge due to the exponentially growing joint action space.
In our analysis, we highlight a critical but often overlooked issue: **extrapolation error**, which arises when unseen state-action pairs are inaccurately assigned unrealistic values, severely affecting performance.
We demonstrate that the success of value factorization methods can be largely attributed to their ability to mitigate this error.
Building on this insight, we introduce multi-step bootstrapping and ensemble techniques to further reduce extrapolation errors, showing that straightforward modifications can lead to substantial performance improvements. Our findings underscore the importance of recognizing extrapolation error in MARL and highlight the potential of exploring simpler methods to advance the field.

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: 5578