Abstract: Concerning the challenges of the diversity of data sources, non-independent and identical distribution of data and the heterogeneity of both computing capabilities and energy consumption of edge devices in Internet of Things (IoT), a computation offloading policy in Mobile Edge Computing (MEC) network that deploys both centralized learning and federated learning was proposed. Firstly, a system model of computation offloading related to both centralized learning and federated learning was built, considering the network transmission delay, computation delay and energy consumption of centralized learning and federated learning models. Then, with the system delay minimization as optimization object, considering the constraints of energy consumption and the training times based on machine learning accuracy, a computation offloading optimization model for machine learning was constructed. After that, the game for this computation offloading was formulated and analyzed. Based on the analysis results, an Energy-Constrained Delay-Greedy (ECDG) algorithm was proposed, which found the optimal solutions for the model via a two-stage policy of greedy decision and energy-constrained decision updating. Compared to the centralized-greedy and Federated Learning with Client Selection (FedCS) algorithms, ECDG algorithm has the lowest average learning delay, which is 1/10 of that in the centralized-greedy algorithm, and 1/5 of that in the FedCS algorithm. The experimental results show that, ECDG algorithms can automatically select the optimal machine learning models by computation offloading so that it can efficiently reduce the average machine learning delay, improve the energy efficiency of edge devices and satisfy the Quality of Service (QoS) requirements of IoT applications.

Key words: Mobile Edge Computing (MEC), computation offloading, machine learning, federated learning, delay sensitive

摘要： 针对物联网（IoT）数据源的多样化、数据的非独立同分布性、边缘设备计算能力和能耗的异构性，提出一种集中学习和联邦学习共存的移动边缘计算（MEC）网络计算迁移策略。首先，建立与集中学习、联邦学习都关联的计算迁移系统模型，考虑了集中学习、联邦学习模型产生的网络传输延迟、计算延迟以及能耗；然后，以系统平均延迟为优化目标、以能耗和基于机器学习准确率的训练次数为限制条件构建面向机器学习的计算迁移优化模型。接着对所述计算迁移进行了博弈分析，并基于分析结果提出一种能量约束的延迟贪婪（ECDG）算法，通过延迟贪婪决策和能量约束决策更新二阶优化来获取模型的优化解。与集中式贪婪算法和面向联邦学习的客户选择（FedCS）算法相比，ECDG算法的平均学习延迟最低，约为集中式贪婪算法的1/10，为FedCS算法的1/5。实验结果表明，ECDG算法能通过计算迁移自动为数据源选择最优的机器学习模型，从而有效降低机器学习的延迟，提高边缘设备的能效，满足IoT应用的服务质量（QoS）要求。

关键词: 移动边缘计算, 计算迁移, 机器学习, 联邦学习, 延迟敏感