Federated Reinforcement Learning for Optimized Spectrum and Mobility Management in 5G NR Networks

Abstract

The exponential growth of connected devices and bandwidth-intensive ap-plications has placed immense demands on modern mobile networks. Fifth-generation (5G) and emerging sixth-generation (6G) networks require real-time, adaptive spectrum and mobility management to meet these expectations. This paper proposes a Federated Reinforcement Learning (FRL) framework for joint handover and spectrum optimization, combining local Deep Q-Networks (DQNs) at distributed gNodeBs with 3GPP-aligned Net-work Data Analytics Function (NWDAF) context prediction and FedAvg-based aggregation. Simulation results using SLAW-modeled user mobility and NR-standard topology demonstrate a 95.2% handover success rate, a 5.2 b/s/Hz spectral efficiency, and a 17.6% improvement in energy efficiency compared to centralized and heuristic baselines. By preserving user privacy and reducing communication overhead, the proposed FRL framework enables scalable, protocol-compliant, and high-performance decision-making for next-generation wireless networks.