Distributed Epigraph Form Multi-Agent Safe Reinforcement Learning

Songyuan Zhang, Oswin So, Mitchell Black, Zachary Serlin, Chuchu Fan

Keywords: Safe multi-agent systems, reinforcement learning, optimal control, epigraph form

TL;DR: We propose EFMARL, a novel MARL algorithm that improves upon the problems faced in the zero constraint threshold setting by extending the epigraph form, a technique to perform constrained optimization, to the CTDE paradigm.

Abstract:

Most existing safe multi-agent reinforcement learning (MARL) algorithms consider the constrained Markov decision process (CMDP) problem, which targets bringing the mean of constraint violation below a user-defined threshold.
However, as observed by existing works albeit for the single-agent case, CMDP algorithms suffer from unstable training when the constraint threshold is zero.
This paper proposes **EFMARL**, a novel MARL algorithm that improves upon the problems faced in the zero constraint threshold setting by extending the *epigraph form*, a technique to perform constrained optimization, to the centralized training and distributed execution (CTDE) paradigm.
We validate our approach in different Multi-Particle Environments and Safe Multi-agent MuJoCo environments with varying numbers of agents. Simulation results show that our algorithm achieves stable training and the best performance while satisfying constraints: it is as safe as the safest baseline that has significant performance loss, and achieves similar performance as baselines that prioritize performance but violate safety constraints.

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: 11688