Joint Data Sampling and Client Scheduling for Over-the-Air Federated Learning
Haichao Wei, Zeke Niu, Bin Lin · 2024
Wireless federated learning (FL) can effectively exploit the rich data distributed at edge devices for network edge intelligence. However, the contradictions between the huge training data and the limited computing capability of devices, as well as between the massive sharing model parameters and the limited communication resources, severely limit the performance of FL at the edge. To overcome these challenges, this paper adopts analog over-the-air computation and jointly considers training sample selection, model pruning, and client scheduling for FL. We first derive an upper bound for the accuracy loss of the aggregation model, considering communication error, model pruning loss, sample selection, and model training cumulative error. A joint problem is formulated based on computation-communication cost and model accuracy loss to balance the training cost and model quality. Then we propose a joint optimization of computation and communication resources algorithm to minimize the objective function. Simulation results demonstrate that this algorithm significantly reduces the overall cost of model aggregation while ensuring model accuracy.