A DRL-Based Server Selection Scheme for IoT Federated Learning in Sparse LEO Satellite Constellations

• Published in: 2024 IEEE 99th Vehicular Technology Conference (VTC2024-Spring) pp. 1 - 5
• Main Authors: Qin, Pengxiang, Xu, Dongyang, Chakraborty, Chinmay, Alfarraj, Osama, Yu, Keping, Guizani, Mohsen
• Format: Conference Proceeding
• Language: English
• Published: IEEE 24.06.2024
• Subjects: deep reinforcement learning; Federated learning; Internet of Things; Low earth orbit satellites; Q-learning; Satellite constellations; Satellites; server selection; Simulation; sparse LEO satellite constellations; Vehicular and wireless technologies
• ISSN: 2577-2465
• DOI: 10.1109/VTC2024-Spring62846.2024.10683018

Federated learning (FL) has emerged in sparse low earth orbit (LEO) satellite constellations as a promising architecture for on-board machine learning (ML) model training, aimed at preserving Internet of Things (IoT) data privacy in specialized and sophisticated tasks. However, the user in FL who spends the longest time in a FL round significantly hinders efficiency. Furthermore, intermittent satellite connectivity, rapidly changing network topologies of sparse LEO satellite constellations and a dearth of information including computation capabilities and positions of satellites greatly obstacle the efficient implementation of FL. To address this challenge, we propose a deep reinforcement learning (DRL)-based server selection scheme for FL in sparse LEO satellite constellations. The optimization problem to minimize the overall FL latency is formulated. A Markov decision process (MDP) is subsequently established and the corresponding double Q-learning agent is trained to make sequential FL server selection decisions to figure it out. Simulation results demonstrate that the proposed scheme reduces latency compared to other server selection schemes.