A Generative Digital Twin Framework for Low-Overhead Multi-Agent Device Scheduling in Wireless Federated Learning

Abstract

Future 6G networks will rely on large numbers of heterogeneous intelligent agents that must coordinate under dynamic wireless conditions. Multi-agent reinforcement learning (MARL) is a promising tool for such coordination, but its training often requires frequent state collection from real devices, causing substantial communication overhead and latency. To address this issue, we propose GenDT, a generative digital twin (DT) framework for low-overhead MARL-based coordination in wireless federated learning (FL). Unlike replicationbased DTs that require continuous synchronisation, GenDT learns the statistical behaviour of wireless FL devices and generates synthetic device states for scheduler training. Specifically, GenDT combines hardware-bias estimation with a conditional variational autoencoder-based wireless channel generator to capture device heterogeneity and channel dynamics. We further propose GenDT-QMIX, integrating GenDT with a QMIX-based multi-agent scheduler to optimise FL device participation under accuracy, energy, and latency objectives. Experiments show that GenDT-QMIX reduces state-collection overhead while maintaining competitive scheduling performance against real-state MARL training and existing RL-based baselines.