This paper investigates the computation offloading problem in a high mobility internet of vehicles (IoVs) environment, aiming to guarantee latency, energy consumption, and payment cost requirements. Both moving and parked vehicles are utilized as fog nodes. Vehicles in high mobility environments need collaborative interactions in a decentralized manner for better network performances, where agent action space grows exponentially with the number of vehicles. The vehicular mobility introduces additional dynamicity in the network, and the learning agent requires a joint cooperative behavior for establishing convergence. The traditional deep reinforcement learning (DRL)-based offloading in IoV ignores other agent’s actions during the training process as an independent learner, which makes a lack of robustness against the high mobility environment. To overcome it, we develop a cooperative three-layer, more generic decentralized vehicle-assisted multi-access edge computing (VMEC) network, where vehicles in associated RSU and neighbor RSUs are in the bottom fog layer, MEC servers are in the middle cloudlet layer, and cloud in the top layer. Then multi-agent DRL-based Hungarian algorithm (MADRLHA) in the bipartite graph maximum matching problem is applied to solve dynamic task offloading in VMEC. Extensive experimental results and comprehensive comparisons are conducted to illustrate the superiority of our proposed method.

中文翻译：

---

用于多接入边缘计算车联网 (IoV) 中任务卸载的基于多代理 DRL 的匈牙利算法 (MADRLHA)


本文研究了高移动性车联网（IoV）环境中的计算卸载问题，旨在保证延迟、能耗和支付成本要求。移动和停放的车辆都被用作雾节点。高移动性环境中的车辆需要以分散的方式进行协作交互，以获得更好的网络性能，其中代理动作空间随着车辆数量呈指数级增长。车辆移动性在网络中引入了额外的动态性，并且学习代理需要联合协作行为来建立收敛。车联网中传统的基于深度强化学习（DRL）的卸载忽略了训练过程中其他智能体作为独立学习者的行为，这使得在高移动性环境下缺乏鲁棒性。为了克服这个问题，我们开发了一个协作的三层、更通用的分散式车辆辅助多路访问边缘计算（VMEC）网络，其中相关 RSU 和相邻 RSU 中的车辆位于底部雾层，MEC 服务器位于中间云层层，云在顶层。然后应用二分图最大匹配问题中基于多智能体DRL的匈牙利算法(MADRLHA)来解决VMEC中的动态任务卸载。进行了大量的实验结果和全面的比较来说明我们提出的方法的优越性。