Research on the Robustness of Cooperative Multi-Agent Deep Reinforcement Learning: A Policy Regularization Method Based on Transfer Learning

Cooperative multi-agent deep reinforcement learning algorithm(c-MADRL)is proven vulnerable to adversarial attacks and its corresponding practical applications are subject to security risks, thus it is crucial for us to improve the robustness of c-MADRL. In this paper, we use a combination of policy...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:Chinese Control Conference S. 5933 - 5939
Hauptverfasser: Zan, Lixia, Zhu, Xiangbin
Format: Tagungsbericht
Sprache:Englisch
Veröffentlicht: Technical Committee on Control Theory, Chinese Association of Automation 28.07.2025
Schlagworte:
Accelerated aging
Deep reinforcement learning
Multi-agent deep reinforcement learning
policy regularization
Robustness
Security
Training
Transfer learning
ISSN:1934-1768
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:Cooperative multi-agent deep reinforcement learning algorithm(c-MADRL)is proven vulnerable to adversarial attacks and its corresponding practical applications are subject to security risks, thus it is crucial for us to improve the robustness of c-MADRL. In this paper, we use a combination of policy regularization and transfer learning to improve the robustness of c-MADRL. Firstly, we use the regularization term based on the worst-case adversarial sample, that is, the adversarial attack that minimizes the expected reward to update the algorithm; Secondly, in order to improve the training efficiency, we employ the policy transfer method, which first performs regularization training on a leader agent, and then transfers the trained policy to other agents; Finally, we verify the performance of the algorithm before and after regularization by adversarial attacks. We have conducted experiments on the multiple-particle environment (MPE), the experimental results show that our methods can improve the robustness of the algorithm and raise defense performance against some adversarial attacks.
ISSN:1934-1768
DOI:10.23919/CCC64809.2025.11179405