Adaptive User Selection and Bandwidth Allocation for Fast Convergence of Federated Learning in Wireless Networks

In resource constrained wireless networks, federated learning (FL) will meet a certain degree of heterogeneity, such as the non independent and identically distributed (non-i.i.d) data, the various communication schemes, the different computation capabilities of massive network users. Under these heterogeneity, the convergence latency of the global model built by FL may increase. Therefore, we firstly propose a novel users’ data importance measurement method and then propose user selection and bandwidth allocation algorithm by maximizing learning efficiency consisting of data importance and latency per round to speed up the convergence of the global model in wireless networks. To solve joint optimization problem with lower complexity, we decouple the above optimization problem into two sub-problems, and then get the best user selection and bandwidth allocation strategy. Our experiments on the Kuzushiji-MNIST and Fashion-MNIST show that the user selection and bandwidth allocation algorithm proposed in this paper can complete the model convergence faster than other traditional algorithms in wireless FL.