LAUSR

This letter considers the beamforming optimization of a heterogeneous intelligent reflective surfaces (IRSs)-aided downlink wireless system. We aim to maximize the achievable sum rate by jointly optimizing the active beamforming at base station (BS) and the passive phase shifts at IRSs, respectively, when maximum transmit power constraint at the BS is met. We propose an unsupervised stratified federated learning algorithm, that allows distributed clients to collaboratively train local models in an unsupervised manner by exploiting internal correlations within the observed channel data while avoiding the effects of heterogeneity. The simulation results demonstrate that the proposed algorithm obtains the same performance compared with the centralized deep learning-based method with much less training data samples, and achieves most of the performance of semidefinite relaxation algorithm with greatly reduced computational complexity.