User Cooperation with Power Control for Federated Learning in CFmMIMO Networks

In synchronous federated learning (SFL), only if all users successfully deliver their local parameters to the central server, the central server can efficiently aggregate a global model. In this case, the user with the worst quality link becomes the bottleneck limiting the training performance of SFL. To cope with this issue for SFL in cell-free massive multiple-input multiple-output (CFmMIMO) network, this paper proposes a joint user cooperation and power control (JUCPC) method. Particularly, the users with good quality links are employed to assist the ones with worse quality links to transfer their parameters to the central server. Then, the power control is employed to coordinate the interference among users to further improve the transmission performance for SFL. To find the joint optimal user cooperation pairing and the power control coefficients, we formulate an optimization problem to minimize the maximum transmission time among users. As the problem is non-convex, we decouple it into two sub-problems. For the first sub-problem, we propose a fairness-based algorithm to select the helping users and also pair them with the ones who required to be assisted. For the second sub-problem, a deep deterministic policy gradient (DDPG)-based algorithm is proposed to find the optimized power control coefficients. Experiments show that the proposed JUCPC method can efficiently reduce the transmission time of the worst user for SFL. Moreover, compared to the traditional optimization method based on convex optimization solver (CVX), our proposed DDPG-based one requires much shorter execution time while achieves approximating result to the CVX-based algorithm. © 2023 IEEE.