第48卷总第523期动态拓扑下的无人机网络计算任务卸载方法职科翔1,李刘杰1,刘晨熙1,李长庚²,彭木根1(1.北京邮电大学信息与通信工程学院,北京100876;2.中南大学电子信息学院,湖南长沙410075)移动通信【摘要】面向动态拓扑下的无人机网络计算任务卸载问题,提出了基于注意力机制和深度强化学习的高性能低复杂度计算任务卸载方法。利用注意力机制动态表征网络中无人机的数量,解决了传统基于深度强化学习的计算任务卸载方法仅适用于固定网络拓扑的难题。在此基础上,提出了预训练和强化学习的级联训练方法,有效地提升了所提方法的收敛速度及性能。仿真结果表明,相比于对比方案,所提算法能显著降低系统的平均时延和丢包率。【关键词】无人机通信;计算任务卸载;注意力机制;强化学习doi:10.3969/j.issn.1006-1010.20240228-0002中图分类号:TN915.41文献标志码:A文章编号:1006-1010(2024)03-0083-07引用格式:职科翔,李刘杰,刘晨熙,等.动态拓扑下的无人机网络计算任务卸载方法[].移动通信,2024,48(3):83-89.ZHIKexiang,LILiujie,LIUChenxi,etal.ComputationTaskOffloadingforUnmannedAerialVehicle-EnabledNetworkswithDynamicTopology[J].MobileCommunications,2024,48(3):83-89.ComputationTaskOffloadingforUnmannedAerialVehicle-EnabledNetworks(1.BejingUniversityofPostsandTelecommunications,SchoolofInformationandCommunicationEngineering,Beijing100876,China;[Abstract]Addressingthecomputationtaskoffloadingchallengeinunmannedaerialvehicle(UAV)-enablednetworkswithdynamictopology,thispaperintroducesahigh-performance,low-complexitycomputationtaskoffloadingapproachleveraginganattentionmechanismanddeepreinforcementlearning.TheattentionmechanismdynamicallyrepresentsthenumberofUAVsinthenetwork,overcomingthelimitationoftraditionaldeepreinforcementlearning-basedoffloadingmethodsthataretailoredonlyforstaticnetworktopologies.Buildinguponthis,acascadedtrainingmethodologycombiningpre-trainingwithreinforcementlearningisproposed,significantlyenhancingtheconvergencespeedandperformanceofthemethod.Simulationresultsdemonstratethattheproposedalgorithmsubstantiallyreducesthesystem'saveragelatencyandpacketlossratecomparedtobenchmarkschemes.[Keywords]unmannedaerialvehi...