联邦学习作为一种分布式机器学习方法，在保证数据不离开边缘设备的情况下，由参数服务器联合客户端共同完成全局模型的训练。由于联邦学习需要在服务器和客户端之间进行频繁的通信，如何对无线通信场景下的联邦学习系统进行性能优化成为了近年来学术界关注的问题。目前，大量工作聚焦于在传统的同步聚合机制下的无线联邦系统的性能优化研究。然而，在系统面临严重的数据异构和设备计算资源异构的情况下，掉队问题会严重降低联邦学习的训练效率。空中计算（Over-the-Air Computation, AirComp）已被证明是一种有效的传输方案，可以提高联邦边缘学习（Federated Edge Learning, FEEL）的效率。然而，现有的采用AirComp方案的FEEL系统通常在每个全局轮次中采用传统的同步聚合机制进行本地模型聚合，这容易受到慢速者问题的影响。本课题提出了一种采用AirComp方案的半异步聚合FEEL机制（PAOTA），以改善在数据和设备异构性显著的情况下FEEL系统的训练效率。考虑到边缘设备模型更新的陈旧性和发散性，本课题通过在每个聚合周期调整边缘设备的上行传输功率，来最小化FEEL全局模型的收敛上界。 仿真实验结果表明，本课题提出的算法实现了接近理想本地SGD的收敛性能。此外，在相同的目标精度下，PAOTA所需的训练时间少于理想本地SGD和通过AirComp的同步FEEL算法所需的训练时间。
The advent of federated learning offers a possibility for training deep learning models under privacy protection.Due to the need for frequent communication between the server and clients in federated learning, how to optimize the performance of federated learning systems in wireless communication scenarios has become a focus of academic research in recent years. Currently, a significant amount of work focuses on performance optimization research for wireless federated systems under traditional synchronous aggregation mechanisms. However, in systems facing severe data heterogeneity and device computational resource heterogeneity, the straggler problem can significantly reduce the training efficiency of federated learning. Over-the-air Computation (AirComp) has been demonstrated as an effective transmission scheme to boost the efficiency of federated edge learning (FEEL). However, existing FEEL systems with AirComp scheme often employ traditional synchronous aggregation mechanisms for local model aggregation in each global round, which suffer from the stragglers issues. In this paper, we propose a semi-asynchronous aggregation FEEL mechanism with AirComp scheme (PAOTA) to improve the training efficiency of the FEEL system in the case of significant heterogeneity in data and devices. Taking the staleness and divergence of model updates from edge devices into consideration, we minimize the convergence upper bound of the FEEL global model by adjusting the uplink transmit power of edge devices at each aggregation period. The simulation results demonstrate that our proposed algorithm achieves convergence performance close to that of the ideal Local SGD. Furthermore, with the same target accuracy, the training time required for PAOTA is less than that of the ideal Local SGD and the synchronous FEEL algorithm via AirComp.
