Joint Optimization of Communications and Federated Learning Over the Air

Federated learning (FL) is an attractive paradigm for making use of rich distributed data while protecting privacy. Nonetheless, nonideal communication links and limited transmission resources may hinder the implementation fast accurate FL. In this paper, we study joint optimization communications FL based on analog aggregation in realistic wireless networks. We first derive closed-form expressions expected convergence rate over air, which theoretically quantify impact Based analytical results, develop a model implementation, allows parameter server to select subset workers determine appropriate power scaling factor. Since practical setting air encounters unobservable parameters, reformulate worker selection allocation using controlled approximation. Finally, efficiently solve resulting mixed-integer programming problem via simple yet optimal finite-set search method by reducing space. Simulation results show that proposed solutions developed channels outperform benchmark method, achieve comparable performance ideal case where implemented noise-free channels.