Generative Multi-Agent Reinforcement Learning for Priority-Aware Task Offloading and Service Caching in Vehicular Edge Networks
Year: 2026
Type: article
Abstract: Vehicular edge computing (VEC) is a key enabler of ultra-reliable and low-latency services in next-generation intelligent transportation systems by bringing computation and service caching closer to vehicles. However, existing joint task offloading and service caching schemes often rely on simplified assumptions, such as single-task vehicle models, static resource availability, or purely reactive control, which fail to capture concurrent multi-task execution, dynamic priority requirements, and stochastic wireless channels in realistic vehicular environments. These limitations lead to inefficient resource utilization and degraded quality of service, particularly for latency-critical applications. In this paper, we propose a Stable Diffusion-based Generative Multi-Agent Reinforcement Learning (SD-GMARL) framework for priority-aware task offloading and service caching in dynamic VEC networks. The proposed framework adopts a hierarchical multi-timescale control architecture, where long-term service caching and resource provisioning decisions are optimized at a slow time scale, while short-term task offloading and scheduling actions are executed at a fast time scale under realistic vehicular mobility and channel conditions. A Stable Generative Diffusion Model (SGDM) is employed to learn the spatiotemporal evolution and uncertainty of vehicular traffic demand and wireless link quality. It produces predictive latent representations that guide cooperative multi-agent actor-critic learning. Simulation results based on standardized vehicular mobility and wireless channel models show that the proposed SD-GMARL framework consistently outperforms representative optimization-based and hierarchical deep reinforcement learning (DRL) baselines in terms of end-to-end latency, energy efficiency, load balancing, and convergence behavior. These results demonstrate the effectiveness of integrating generative foresight with hierarchical multi-agent learning for predictive and priority-aware resource management in highly dynamic VEC systems.
Language: English
Topic: IoT and Edge/Fog Computing
Subfield: Computer Networks and Communications
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $2,645
OpenAlex ID: W7204143402
