Flexible Training and Uploading Strategy for Asynchronous Federated Learning in Dynamic Environments

Federated learning is a fast-developing distributed learning scheme with promising applications in vertical domains such as industrial automation and connected automated driving. The heterogeneity of devices in data distribution, communication, and computation, when deployed in dynamic environments typically with wireless communication, poses challenges to traditional federated learning solutions, where successful learning depends on balanced contribution from participants. In this paper, we propose a flexible communication strategy for devices in asynchronous federated learning, which adapts the training and uploading actions based on the condition of the communication link. We propose a novel method of computing aggregation weight based on model distances and number of local optimizations, to control errors introduced in asynchronous aggregation while maximizing learning speed. We prove the convergence of the learning tasks analytically under the new scheme. The improved performance is rooted in the increased number of optimizations during training, which grows by 12% through opportunistically condensing model uploading during good link condition periods. By facilitating timely communication between devices and server, combined with the novel aggregation weight design, our method reduces the communication resources in dynamic environments by at least 5% while even slightly increasing the learning accuracy.