AI-Driven Predictive Handover in Ultra-Dense 6G Networks: A Federated Graph Reinforcement Learning Approach

Fehri, Hamed; MonemiZadeh, Mostafa

doi:10.22034/cas.2026.245730

AI-Driven Predictive Handover in Ultra-Dense 6G Networks: A Federated Graph Reinforcement Learning Approach

Document Type : Original Article

Authors

Hamed Fehri ¹

Mostafa MonemiZadeh ²

¹ Department of Electrical Engineering, Faculty of Engineering, University of Zabol, Zabol, Iran

² Department of Electrical Engineering, University of Neyshabur, Neyshabur, Iran

https://doi.org/10.22034/cas.2026.245730

Abstract

Traditional handover (HO) methods encounter critical challenges in satisfying the Tbit/s throughput and sub-millisecond latency requirements of ultra-dense 6G networks. Inability to achieve ultra-reliable and low-latency communication (URLLC), excessive overhead of signaling and ping-pong effects are critical issues arising from applying traditional 3GPP-compliant HO methods in ultra-dense networks (UDNs) with device densities exceeding 700 UEs/km². To tackle these limitations, this paper proposes a new AI-based predictive HO architecture for pre-planning HO timing and target BS in advance up to 500 ms. Our proposed architecture is a hybrid one based on federated graph reinforcement learning (FGRL) and can provide an intelligent and proactive mobility management while preserving privacy in 6G networks. In the proposed FGRL method, federated learning addresses the privacy concerns, graph neural networks model the spatial relationships between BSs, and Multi Agent Deep Q Networks (MADQN) provide distributed decision making. To assess the performance of the proposed method, extensive simulations are conducted showing that the proposed method significantly improves HO latency, ping-pong rates, and throughput compared to 3GPP-standard-compliant A3/A6 event-based algorithms. We also compare centralized, distributed, and federated designs and demonstrate that the proposed federated design achieves competitive performance to the centralized design while preserving user privacy and considerably decreasing uplink overhead.

Graphical Abstract