Maintaining Coordination in Multi-Agent Reinforcement Learning Under Pairwise Asynchronous Link Failures Through Cached-Message Reuse
Applied Sciences · 07.11.2026 · 16 (21) · 11031
DOI: 10.3390/app162111031 ↗Kurz gefasst
Fachartikel in der Zeitschrift Applied Sciences, online erschienen am 07.11.2026. Verfasst von Peiying Zhang und 5 weiteren Autor:innen. Die Arbeit nennt 40 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: National Natural Science Foundation of China, Natural Science Foundation of Shandong Province, the Shandong Provincial Undergraduate Teaching Reform Research Project.
Abstract (Originalsprache)
Cooperative multi-agent reinforcement learning (MARL) relies on communication when local observations are incomplete. Pairwise asynchronous link failures interrupt different directed links at different times and remove part of the information used for coordination. This paper studies whether agents can maintain coordination by reusing the latest message received from each unavailable neighbor. We propose Temporal Graph-based Neighbor Aggregation with Age-Penalized Communication (TGN-APC). Each agent stores one message per neighbor and limits reuse by message age. TGN-APC reuses a valid cached message and subtracts an age penalty from its attention score. We evaluate the method on the Vectorized Multi-Agent Simulator Navigation task with ten independent training seeds. Cache reuse significantly improves return and team completion over direct one-hop communication under short and long failure bursts. An age penalty coefficient of 0.10 does not significantly improve on cache reuse without an age term. Matched validation-based coefficient selection shows no significant difference between score penalization and exponential content scaling after multiple-testing correction. A two-hop current-message variant gives the highest long-burst return, but it uses about 83% more active transmissions than TGN-APC. These results support cached-message reuse for the tested task and failure model. They do not establish a separate performance gain from the selected age penalty.
Förderung
- National Natural Science Foundation of China
- Natural Science Foundation of Shandong Province
- the Shandong Provincial Undergraduate Teaching Reform Research Project
Zitieren
APA
Zhang, P., Yu, J., Tan, L., Kostromitin, K. I., Wang, J., & He, S. (2026). Maintaining Coordination in Multi-Agent Reinforcement Learning Under Pairwise Asynchronous Link Failures Through Cached-Message Reuse. Applied Sciences, 16(21), 11031. https://doi.org/10.3390/app162111031
Vancouver
Zhang P, Yu J, Tan L, Kostromitin KI, Wang J, He S. Maintaining Coordination in Multi-Agent Reinforcement Learning Under Pairwise Asynchronous Link Failures Through Cached-Message Reuse. Applied Sciences. 2026;16(21):11031. doi:10.3390/app162111031
BibTeX
@article{Zhang2026,
author = {Zhang, Peiying and Yu, Jun and Tan, Lizhuang and Kostromitin, Konstantin Igorevich and Wang, Jian and He, Shuqing},
title = {Maintaining Coordination in Multi-Agent Reinforcement Learning Under Pairwise Asynchronous Link Failures Through Cached-Message Reuse},
journal = {Applied Sciences},
year = {2026},
volume = {16},
number = {21},
pages = {11031},
doi = {10.3390/app162111031}
}
EUVIMED