A Vehicular Networking Routing Protocol Based on Improved Reinforcement Learning

In response to the challenge of achieving high reliability and adaptability in packet routing protocols within complex vehicular networking environments and the difficulty of training reinforcement-learning frameworks on a large number of real devices, we propose an enhanced reinforcement-learning packet protocol with multi-agent integration based on the existing three-tier edge intelligence architecture (user layer, edge collaboration layer, and cloud computing layer). The multi-hop transmission of multiple packets was modeled as a Markov process, and improvement measures were proposed for a reinforcement learning framework and corresponding effective training methods. And propose a solution to deploy a vehicle network simulation platform in the cloud to complete the pre-training process of reinforcement learning frameworks. Then, the trained reinforcement learning model was deployed on edge servers to guide the direction of the packet transmission. The experimental results demonstrate that, compared to existing vehicular networking routing protocols, the proposed protocol achieves at least a $30 \%$ improvement in message transmission delay and an $18 \%$ improvement in reception rate.