Abstract: In vehicular edge computing (VEC) environments, the high mobility and uneven distribution of vehicles often result in cross-domain task offloading failures and load imbalance among roadside units (RSUs), thereby degrading overall system performance. Existing approaches primarily emphasize either load balancing or single-step offloading decisions, while insufficiently addressing communication uncertainties caused by vehicle mobility and the requirement for flexible task scheduling across RSUs, which limits offloading efficiency. To overcome these limitations, a trajectory-aware multi-agent reinforcement learning–based cooperative offloading method is introduced. A PatchTST-based multi-step vehicle trajectory prediction module is integrated to provide foresighted mobility awareness for task scheduling. A vehicle–edge–cloud collaborative offloading framework is further developed, in which vehicles and RSUs are modeled as agent clusters. A centralized training and decentralized execution (CTDE) paradigm enable joint decision-making, and a partially reward-decoupled multi-agent policy optimization mechanism is designed to improve cooperation efficiency and system stability. Extensive simulations under varying task densities, delay constraints, and network topologies show that the proposed method consistently outperforms representative baselines such as IPPO, MADDPG, and MAPPO, achieving an average improvement of 36.1% in task completion rate and a reduction of 37.6% in task completion delay. These results demonstrate the effectiveness, robustness, and practical potential of the proposed approach in dynamic and complex VEC environments.

Key words: Vehicular Edge Computing, Multi-Agent Reinforcement Learning(MARL), trajectory prediction, task offloading, collaborative task scheduling

摘要： 在车辆边缘计算（Vehicular Edge Computing, VEC）环境中，车辆的高速移动性和非均匀分布常导致跨域任务卸载失败和路侧单元（Roadside Units, RSUs）负载失衡，从而严重影响系统性能。现有研究多集中于负载均衡或仅考虑单步卸载决策，忽略了车辆移动性带来的通信不确定性以及任务在跨RSU环境下的灵活调度需求，导致卸载效率受限。为此，本研究提出了一种融合轨迹预测的多智能体强化学习协同卸载方法。该方法首先引入基于PatchTST的多步车辆轨迹预测模块，实现前瞻性的移动感知以指导任务调度；随后构建车-边-云协同卸载框架，将车辆与RSU建模为智能体群体，采用集中训练、分散执行的强化学习范式进行联合决策，并设计了部分奖励解耦的多智能体策略优化方法以提升协作效率与系统稳定性。在不同任务密度、时延约束及网络拓扑的仿真测试中，所提方法相较IPPO、MADDPG及MAPPO等先进基线，任务完成率平均提升约36.1%，任务完成时延平均降低约37.6%，显著优于现有方法。实验结果验证该方法在动态复杂VEC环境下的高效性、鲁棒性与应用潜力。

关键词: 车辆边缘计算, 多智能体强化学习, 轨迹预测, 任务卸载, 协同任务调度