To address the challenging issues in task offloading and resource allocation within the Internet of Vehicles (IoV) environments, such as multi-objective task offloading, hybrid action space decision-making, and low training efficiency, a hybrid asynchronous advantage actor-critic (H-A3C) reinforcement learning algorithm was proposed by integrating the hybrid action representation (HyAR) framework with an improved asynchronous training mechanism. The optimization problem of multi-objective task offloading and resource allocation in IoV was investigated. In this approach, a parameterized action Markov decision process (PAMDP) was constructed to uniformly model the coupling relationship between discrete offloading target selection and continuous resource allocation. A conditional variational autoencoder (CVAE) was designed to achieve a latent space representation of hybrid actions, and a time-decaying gradient aggregation mechanism was introduced to enhance the stability of asynchronous training. The results demonstrate that H-A3C significantly outperforms relevant baseline algorithms in terms of convergence speed and stability. The H-A3C algorithm stabilizes after approximately 3 500 training episodes, which is about 40% faster than the standard A3C algorithm. In terms of cumulative system reward, its final return is improved by 25% compared to the PADDPG algorithm and by 35% compared to the standard A3C. Under high vehicle density conditions (more than 110 vehicles), by precisely adapting the offloading ratio, H-A3C increases edge resource utilization by about 18% and reduces reliance on cloud resources by 30%, leading to superior system returns. In computationally intensive task scenarios, its task success rate is 15% and 22% higher than that of PADDPG and standard A3C, respectively. Moreover, it can dynamically adjust optimization weights, reducing the total processing latency by approximately 20% and leasing costs by about 15%. thereby achieving a better latency-cost trade-off and effectively lowering the overall processing time and rental expenses.
LIU Qiu-ming, GUO Xing-xing, CHENG Zhi-kang, ZHOU Xin-cheng, XIAO He.
Multi-task Offloading and Resource Allocation for Internet of Vehicles[J].
Science Technology and Engineering, 2026, 26(25): 10909-10920 DOI:10.12404/j.issn.1671-1815.2506464
随着智能交通系统的持续发展,车联网(Internet of Vehicles,IoV)应用正呈现出多样化的演进趋势,包括自动驾驶、智能导航及增强现实等数据密集型服务[1]。此类应用通常依赖于处理海量车载传感器生成的时序数据,具有计算密集与时延敏感的特性,其资源需求常超出单车计算能力。传统云计算架构中车辆与远程云服务器间存在显著传输时延,难以满足高时敏任务需求。为此,移动边缘计算(mobile edge computing,MEC)技术被引入车联网体系。它可将计算、存储及通信能力下沉至网络边缘,支持车辆与周围环境(vehicle-to-everything,V2X)进行实时信息交换,为IoV任务提供低延迟、高响应的卸载服务。
感知车辆会随机生成具有不同计算资源需求的任务,并向邻近节点[计算车辆或路侧单元(road side unit,RSU)]广播卸载请求。响应节点需满足最大通信半径约束:计算车辆的响应范围为车车通信半径RV2C,MEC服务器的响应范围为通信半径RV2M。在最大通信半径内的RSU及拥有空闲资源的计算车辆会响应其ID、位置距离d和可用计算资源信息(包括可用的带宽与计算能力)。假设生成的任务的数据具有细粒度特性,可被动态拆分为任意大小的并行子任务集。将在时隙t的计算总任务定义为三元组Tsk(t)(S,ρ,Tmax),其中S为总任务的待处理数据量(单位:bit),ρ为任务计算密度(GFLOPS/bit),Tmax表示最大容忍时延(满足条件|Δt|>>Tmax)。系统的任务决策机制支持多目标并行卸载,定义卸载目标类型集合Κ∈{L,Csel,Msel,N}。任务的离散卸载决策(卸载目标选择)定义为ad=Kactive,其中Kactive⊆Κ,Kactive≠∅:存储激活的目标类型的子集;Csel⊆Vcomp:存储选定的计算车辆ID子集;Msel⊆Μ:存储选定的MEC服务器ID子集。
任务卸载的总时延Tk由数据传输时延Tktrans与计算处理时延Tkcomp两部分构成。由于任务支持多目标并行卸载,需分别计算各卸载目标子任务的时延,并考虑其并行执行特性。为精确刻画车联网动态无线环境对卸载性能的影响,引入了包含视距(line of sight,LOS)传播状态、多普勒频移及动态覆盖范围的复合信道模型。
MaZhonggui, LiZhuo, LiangYanpeng. A survey on integrated sensing, communication, and computing in internet of autonomous vehicles: review and prospects[J]. Chinese Journal of Engineering, 2023, 45(1): 137-149.
[3]
ChengZ, MinM, LiwangM, et al. Multiagent DDPG-based joint task partitioning and power control in fog computing networks[J]. IEEE Internet of Things Journal, 2021, 9(1): 104-116.
[4]
XiaoL, LuX, XuT, et al. Reinforcement learning-based mobile offloading for edge computing against jamming and interference[J]. IEEE Transactions on Communications, 2020, 68(10): 6114-6126.
LiQiang, YiJinhui, DuTingting, et al. Dependent task offloading and resource allocation based on A3C in mobile edge computing[J]. Computer Engineering, 2023, 49(6): 42-52.
[7]
GuoH, LiuJ, RenJ, et al. Intelligent task offloading in vehicular edge computing networks[J]. IEEE Wireless Communications, 2020, 27(4): 126-132.
ZhaoHaitao, ZhangTangwei, ChenYue, et al. DQN-based task offloading algorithm for vehicular edge networks[J]. Journal on Communications, 2020, 41(10): 172-178.
[10]
LiM, YuF R, SiP, et al. Resource optimization for delay-tolerant data in blockchain-enabled IoT with edge computing: a deep reinforcement learning approach[J]. IEEE Internet of Things Journal, 2020, 7(10): 9399-9412.
ZhaoChanchan, GuoXiaomin, HaiXiaowei, et al. Joint optimization strategy for task offloading and resource allocation in cache-assisted mobile edge computing[J]. Science Technology and Engineering, 2023, 23(9): 3812-3819.
[13]
ChenJ, XingH, XiaoZ, et al. A DRL agent for jointly optimizing computation offloading and resource allocation in MEC[J]. IEEE Internet of Things Journal, 2021, 8(24): 17508-17524.
[14]
LiuY, CuiQ, ZhangJ, et al. An actor-critic deep reinforcement learning based computation offloading for three-tier mobile computing networks[C]// 2019 11th International Conference on Wireless Communications and Signal Processing (WCSP). Xi’an: IEEE, 2019: 1-6.
RuanYingjun, HouZequn, QianFanyue, et al. Optimization of the operation of distributed energy system based on deep reinforcement learning[J]. Science Technology and Engineering, 2022, 22(17): 7021-7030.
ZhouHengheng, GaoSong, WangPengwei, et al. Intelligent vehicles behavior decision-making based on deep reinforcement learning[J]. Science Technology and Engineering, 2024, 24(12): 5194-5203.
[19]
HausknechtM, StoneP. Deep reinforcement learning in parameterized action space[J]. arXiv prePrint, arXiv: 1511. 04143, 2015.
[20]
AleL, ZhangN, FangX, et al. Delay-aware and energy-efficient computation offloading in mobile-edge computing using deep reinforcement learning[J]. IEEE Transactions on Cognitive Communications and Networking, 2021, 7(3): 881-892.
[21]
XiongJ, WangQ, YangZ, et al. Parametrized deep Q-networks learning: reinforcement learning with discrete-continuous hybrid action space[J]. arXiv prePrint, arXiv: 1810. 06394, 2018.
[22]
FuH, TangH, HaoJ, et al. Deep multi-agent reinforcement learning with discrete-continuous hybrid action spaces[J]. arXiv prePrint, arXiv: 1903. 04959, 2019.
[23]
RazaS J A, LinM. Constructive policy: reinforcement learning approach for connected multi-agent systems[C]// 2019 IEEE 15th International Conference on Automation Science and Engineering (CASE). Vancouver, BC: IEEE, 2019: 257-262.
[24]
ChengZ, MinM, GaoZ, et al. Joint task offloading and resource allocation for mobile edge computing in ultra-dense network[C]// GLOBECOM 2020-2020 IEEE Global Communications Conference. Taipei, China: IEEE, 2020: 1-6.