Optimized Scheduling Transmissions for Wireless Powered Federated Learning Networks

Abstract

We have developed a resource allocation scheme that minimizes the training process of federated machine models in the wireless powered communication networks. The new resource sharing method allows energy harvesting (EH) clients (EHCs) to train their local models for extended periods that overlap with data transmissions of other EHCs. Training latency minimization leads to mixed integer non-convex problem, which is tackled by exploiting the sensitivity properties of the corresponding Lagrange multipliers. If the local training models at all EHCs use equal size datasets, the optimal transmission order is in the decreasing order of the EHC-base station channels gains. The proposed resource allocations significantly reduce the training latency compared to the state-of-the-art benchmark schemes.