Deep Reinforcement Learning-Based Resource Allocation for UAV-Enabled Federated Edge Learning

作     者：Tianze Liu Tiankui Zhang Jonathan Loo Yapeng Wang Tianze Liu;Tiankui Zhang;Jonathan Loo;Yapeng Wang

作者机构：School of Information and Communication EngineeringBeijing University of Posts and TelecommunicationsBeijing 100876China School of Computing and EngineeringUniversity of West LondonLondon W55RFUK Faculty of Applied SciencesMacao Polytechnic UniversityMacao 999078China 

出 版 物：《Journal of Communications and Information Networks》 (通信与信息网络学报（英文）)

年 卷 期：2023年第8卷第1期

页      面：1-12页

核心收录：

学科分类：0402[教育学-心理学(可授教育学、理学学位）] 0303[法学-社会学] 0710[理学-生物学] 08[工学] 081104[工学-模式识别与智能系统] 0804[工学-仪器科学与技术] 080402[工学-测试计量技术及仪器] 0811[工学-控制科学与工程] 0812[工学-计算机科学与技术（可授工学、理学学位）] 0713[理学-生态学] 

基　　金：supported by Beijing Natural Science Foundation under Grant 4222010. 

主　　题：DRL FL MEC UAV 

摘      要：The resource allocation of the federated learning(FL)for unmanned aerial vehicle(UAV)swarm systems are investigated.The UAV swarms based on FL realize the artificial intelligence(AI)applications by means of distributed training on the basis of ensuring the security of private data.However,the direct application of the FL in UAV swarms will incur high overhead.Therefore,in this article,we consider the resource allocation problem in FL for UAV swarms.To avoid the high communication overhead between UAVs and the central server,we proposed an FL framework for UAV swarms based on mobile edge computing(MEC)in which model aggregation is migrated to edge servers.In the proposed framework,the total cost of the FL is defined as the weighted sum of the total delay of UAV swarms to complete the FL and system energy consumption.In order to minimize the total cost of FL,we propose a resource allocation algorithm for joint optimization of computing resources and multi-UAV association based on deep reinforcement learning(DRL).The simulation result shows that:(1)compared with the benchmark algorithm,the proposed algorithm can effectively reduce the total cost of FL;(2)the proposed algorithm can realize the trade-off between task completion delay and system energy consumption through weight changes.