Priority-Aware Multi-Agent Q-Learning Routing for Decentralized Wireless Federated Learning
Ref: CISTER-TR-250902 Publication Date: 8, Nov, 2025
Priority-Aware Multi-Agent Q-Learning Routing for Decentralized Wireless Federated Learning
Ref: CISTER-TR-250902 Publication Date: 8, Nov, 2025Abstract:
Decentralized federated learning (DFL) improves flexibility and scalability via peer-to-peer communication in mobile wireless networks. To tackle limited node resources and unreliable links in multi-hop DFL, we propose Priority-aware Multi-agent Q-learning Routing (PMQR) to enhance model transmission reliability and reduce congestion. Each agent optimizes routing for a client pair by maximizing its reward and competing for resources. Using a source-aware penalty and priority mechanism, agents avoid congestion and choose better routes. Experiments show PMQR boosts test accuracy by 23% over baseline DFL without congestion control, and adds 8% training accuracy improvement with the proposed mechanisms.
Xiaoyu He
Kai LiAccepted in ACM Workshop on Mobility in the Evolving Internet Architecture (MobiArch) (MobiArch).
Hong Kong, China.
Record Date: 9, Sep, 2025
Short links for this page: www.cister-labs.pt/docs/cister_tr_250902 www.cister-labs.pt/docs/1971