Toward Energy-Efficient Multiple IRSs: Federated Learning-Based Configuration Optimization

Intelligent reflecting surface (IRS) can enhance the capacity and cost-effectiveness in future wireless networks substantially. However, the configuration optimization of IRS in an energy-efficient way is still a challenging work. In this paper, we propose a solution to the problem of maximizing the total throughput of a multiple IRSs assisted multi-user communication system. A federated deep learning (FDL) based algorithm is designed to obtain the optimal reflection configurations of all IRSs in parallel, where the model parameters are transmitted instead of the dataset itself as in deep learning (DL). Specifically, a deep neural network (DNN) is formulated to fit the coupling relationship between the coordinate information of users and the optimal reflecting vector of IRS. Meanwhile, the analysis of transmission and computation overhead is performed to establish an accurate energy consumption model. For performance evaluation, we conduct a series of simulations to verify the effectiveness of the FDL framework. The simulation results demonstrate that the test accuracy of the FDL framework is as high as 95.22% with only 1/36 of the transmission energy consumption compared with the DL. Moreover, the total throughput can achieve 93% of the theoretical performance.