Safe Multi-agent Reinforcement Learning with Protection Motivation Theory

Xin He, Hongwei Ge, Chunguo Wu, Jincheng Yu

Keywords: Safety, Multi-agent Reinforcement Learning, Protection Motivation Theory

Abstract: A challenging problem for implementing multi-agent reinforcement learning (MARL) in real-world applications is ensuring the safety of cooperative strategies. According to the Protection Motivation Theory (PMT), threat appraisals result in negative emotions and elicit protective behaviors, which are instrumental for coping with security threats. Drawing inspiration from the PMT, we focus on two discrete emotions--fear and regret--to evaluate threat severity and facilitate multiple agents to learn protective behaviors. These can promote cooperative decision-making with fewer safety violations. Specifically, we propose two safety guarantee methods with PMT: fear for safety guarantee (F4SG) and regret for safety guarantee (R4SG), utilizing the active inference technique to model the emotions of fear and regret separately. The threat severity evaluated by these emotions influences the state value and the executed action respectively, which avoids the potential threat of visiting certain states or taking certain actions. Experimental results demonstrate that our proposed methods are safer and more efficient than state-of-the-art baselines on challenging tasks in safe MARL benchmarks.

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.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

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: 42