With the development of intelligent transportation systems, vehicular applications demonstrate diverse characteristics, including computation-intensive processing and stringent latency requirements. Traditional computation offloading strategies struggle to cope with the highly dynamic, multi-node, and multi-task concurrent vehicular network environment and generally overlook the risk of cross-zone communication failures caused by high-speed mobility. To address this issue, this paper designs a computation offloading algorithm based on multi-agent reinforcement learning. This method comprehensively considers four heterogeneous features including queue load, communication links, task attributes, and computing resources, establishes a multi-layer collaborative computing architecture integrating task migration and result return mechanisms, and further constructs an optimization model aimed at minimizing the weighted sum of latency and energy consumption. This model is formalized as a multi-agent Markov decision process, and an improved Multi-Agent Proximal Policy Optimization(MAPPO)-based MATPPO-T algorithm is designed to solve it, achieving one-step joint optimization of task offloading, resource allocation, and task result migration. Experimental results demonstrate that the proposed method reduces the total system cost by approximately 22% on average compared to benchmark algorithms such as MAPPO and PPO, while consistently maintaining the lowest offloading overhead and fastest convergence speed, validating its robustness and scalability in dynamic vehicular edge networks.
Liu et al. (Fri,) studied this question.