jTOLP-MADRL: A MADRL-Based Joint Optimization Algorithm of Task Offloading Location and Proportion for Latency-Sensitive Tasks in Vehicle Edge Computing Network

In Vehicle Edge Computing Network (VECN), task offloading is a key technique to provide the satisfactory quality of service (QoS) for latency-sensitive tasks. However, the diversity of computational resources in edge nodes (i.e., RSU and idle vehicles) and the mobility of vehicles present significant challenges to task offloading. Hence, to address these challenges, we propose an offloading scheme that jointly allocates RSU nodes (including MEC servers) and idle service vehicle resources in this paper. We first prioritize these tasks based on their maximum tolerable latency and design a utility function to capture the executing cost for latency-sensitive tasks. Then, we propose a joint optimization algorithm of task offloading location and proportion based on Multi-agent Deep Reinforcement Learning (jTOLP-MADRL algorithm) for latency-sensitive tasks in VECN, which consists of two sub-algorithms: the Offloading Location Selection (OLS) algorithm and the Offloading Proportion Allocation (OPA) algorithm. Additionally, we design a Convolutional Recurrent Actor-Critic Network (CRACN) to enhance the learning efficiency of the OLS algorithm. Finally, we indicate our algorithm is effective based on simulation results. Compared with the other benchmark algorithms, jTOLP-MADRL can significantly reduce latency and enhance system utility.