A Contextual Multi-Armed Bandit-based Client Selection Algorithm in Digital Twin-Empowered Federated Learning

• 1、School of Computer Science and Technology, Beijing Jiaotong University, Beijing 100044

Abstract：To address the high computation and communication cost as well as the limited resources of physical clients in Federated Learning (FL), this paper proposes a cost-efficient Digital Twin (DT)-empowered federated learning framework. The proposed framework goes beyond existing studies where digital twins are primarily used for monitoring and decision support, by directly threating them as selectable training executors. Specifically, each edge–cloud pair is modeled as a Training Unit composed of a Physical Client (PC) and its corresponding Digital Twin. Each training unit has two execution mode, the PC mode provides high-fidelity but resource-intensive local training, while the DT mode offers a low-cost alternative despite data synchronization discrepancies. During the client selection process in federated learning, both the training unit and the execution mode within the unit are selected simultaneously.Furthermore, to address the uncertainty in training utility, client selection and execution-mode switching are jointly formulated as a Contextual Multi-Armed Bandit (CMAB) decision problem, where the update utilities of the two modes are learned separately from historical feedback. Experiments conducted on a real-world traffic flow dataset show that proposed framework maintains comparable model convergence while reducing time overhead by approximately 10.7\%–24.1\% and energy consumption by 10.9\%–23.2\%, achieving an effective balance between training performance and resource efficiency.

Keywords： Federated Learning Digital Twin Contextual Multi-Armed Bandit Client Selection Resource optimization