Joint client selection and resource allocation for federated edge learning with imperfect CSI

Federated edge learning (FEL) has become a key technology due to its privacy protection for clients. Since during the FEL process, there always exists the parameter passing between edge clients and server under an open communication environment, the learning performance depends heavily on the wireless channel conditions. In this paper, we investigate the performance optimization of FEL system in a practical Internet of Things (IoT) scenario where the channel state information (CSI) is imperfect. A non-convex joint optimization problem for client selection and resource allocation is first built to balance the total energy consumption and learning accuracy of the FEL system. Then for solving the built optimization problem with mixed integer properties, two subproblems are derived by relaxing and dividing. For the resource allocation subproblem, a resource allocation algorithm based off-policy optimization (RAOPO) is proposed to obtain the resource allocation scheme. Based on the resource allocation, an energy-efficient and low-latency client selection algorithm (ELCS) is further designed for improving the performance. The extensive simulations verify that, when considering the imperfect CSI, our proposed ELCS can ensure the learning accuracy and system stability with a low energy cost, which supports the fast development of FEL.