Stochastic Client Scheduling with Dynamic SINR Thresholds for Fast Federated Learning

Federated edge learning is often used in delay-sensitive big data machine learning. However, in cellular networks with limited communication resources, efficient scheduling of large groups of clients has become vital. There is evidence that the impact of the number of participating clients in initial communication rounds is less significant than that of subsequent communication rounds in global model aggregations. Inspired by this, in this paper, we propose a Stochastic Client Scheduling Strategy with Dynamic SINR Thresholds (STUART). In particular, we apply a higher SINR threshold for initial communication rounds in the base station to prevent low-rate clients from participating in the model aggregation, which can effectively reduce local model upload time. Thereafter, the SINR threshold is gradually lowered for the following communication rounds to allow more clients to take part in, hence improving convergence performance. With a convergence constraint, the optimal client scheduling problem is formulated and solved using the genetic algorithm. Finally, we verify the effectiveness of the proposed algorithm through extensive experiments. The results reveal that compared with the fixed threshold method, the STUART algorithm can reduce communication time by more than 1/3 while marginally degrading the convergence performance.