Energy-Efficient Resource Allocation in IRS-Assisted NOMA Networks: A Hierarchical Federated Learning Approach

Abstract

To tackle signal blockage and energy constraints in next-generation networks, this letter proposes a mobile intelligent reflecting surface (IRS)-assisted multi-cell non-orthogonal multiple access (NOMA) network. An optimization problem is formulated to maximize system energy efficiency by jointly optimizing the IRS trajectory, phase shift configuration, and user power allocation. To mitigate the signaling overhead and privacy risks associated with traditional centralized algorithms, a hierarchical federated learning (HFL) framework is proposed. Specifically, users conduct local training by deep Q-networks (DQN), UAVs perform intra-cell model aggregation, and a central server conducts inter-cell global aggregation. By incorporating a weighted aggregation strategy and a hierarchical architecture, the proposed algorithm further reduces model aggregation errors while enabling environmental exploration and model parameter exchange. Simulation results demonstrate that the HFL-DQN algorithm significantly enhances the system’s energy efficiency, while achieving faster convergence than the independent training framework.