A MAPPO-Based Dynamic Multi-Channel Access Method for Wireless Federated Learning
Xin Li, Yitong Li · 2025
Recently, Federated Learning (FL) has emerged as a highly efficient distributed learning paradigm, enabling collaborative model training by sharing model parameters among multiple clients rather than raw data. However, clients with varying degrees of non-independent and non-identically distributed (non-lID) data exhibit divergent local update norms, which can introduce imbalances during model training and lead to disparities in model accuracy. To address this challenge, we leverage the local update norm as a key metric to differentiate clients. Specifically, we propose a Multi-Agent Proximal Policy Optimization (MAPPO) algorithm to tackle the dynamic multi-channel access problem for clients with heterogeneous local update norms in wireless networks. Experimental results demonstrate that the proposed MAPPO algorithm, which incorporates local update norm feedback, significantly accelerates model convergence and enhances model performance in scenario involving non-lID data distributions among clients.