Reinforcement Learning-Based Task Offloading For Air-Ground Integrated Vehicular Edge Computing

As an extension of the Internet of Things(IoT),Internet of Vehicles(IoV)cannot only provide an efficient and comfortable driving experience with people,but also promote the diversification of transportation services.In order to obtain higher reliability and lower latency V2X services,the widely deployed Mobile Edge Computing(MEC)is considered to be a promising solution,which not only effectively reduces the delay and energy consumption caused by lacking computing power of the mobile terminal,but also improves the Quality of Service(QoS)and Quality of Experience(QoE).However,relying only on ground infrastructure cannot provide all-vehicle communication device with ultra-reliable low-latency communication(URLLC)constraints.At present,using aerial platforms as base stations,combined with ground network communication,is one of the most promising solutions for large-capacity communications and emergency communications.Such as tethered balloons,airplanes,and unmanned aerial vehicles(UAVs)that are between satellite and ground,which can increase communication coverage effectively and provide wireless narrowband and broadband services.However,the optimization strategy of the existing research about air-ground integrated vehicle edge computing is limited to single or average QoS performance indicators,which cannot meet the individual needs or subjective experience of user vehicles(UVs).In addition,the mobile characteristics of UAVs and user vehicles pose new challenges for online optimization task offloading.Therefore,it is necessary to design a reasonable and efficient online task offload optimization scheme to ensure the high reliability and low latency of 5G communication while ensuring the QoE performance of user vehicles.Combined with the Air-Ground Integrated Vehicular Edge Computing(AGI-VEC)scenario,this paper establishes intent modeling taking both subjective and objective performance metrics into consideration.Secondly,considering the limited computing capability,prohibitive signaling overheads,and privacy as well as security concerns,we propose a Learning-based Intent-aware Upper Confidence Bound(LIUCB)algorithm based on the Multi-Armed Bandit(MAB)model and the Lyapunov optimization.LIUCB can achieve three-dimension intent awareness including QoE awareness,URLLC awareness,and trajectory similarity awareness.Simulation results demonstrate that LIUCB significantly outperforms three benchmark algorithms in terms of QoE,end-to-end latency,queuing delay,throughput,and times of task offloading failure.