A Digital Twin-based multi-objective optimized task offloading and scheduling scheme for vehicular edge networks

Traditional research on vehicular edge computing often assumes that the requested and processed task types are the same or that the edge servers have identical computing resources, ignoring the heterogeneity of task types in mobile vehicles and the services provided by edge servers. Meanwhile, the complexity of the vehicular edge environment and the large amount of real-time data required by DRL are often ignored when using Deep Reinforcement Learning (DRL) to process the vehicular edge tasks; Furthermore, traditional offloading and scheduling models are usually based on idealized models with deterministic task quantities and a single objective (such as latency or energy consumption). This paper proposes a Digital Twin(DT)-based multi-objective optimized task offloading and scheduling scheme for vehicular edge networks to address these issues. To address the complexity of vehicular edge environments and the need for a large amount of real-time data for DRL, this paper designs a DT-assisted vehicular edge environment; To tackle the problem of task heterogeneity in mobile vehicles and edge server service differentiation, a computation model based on Deep Neural Networks (DNN) partitioning and an early exit mechanism is proposed, which leverages the resources of mobile vehicles and edge servers to reduce the time and energy consumption of DNN tasks during the computation process. For the uncertain task quantity of DNN tasks, a schedule model based on the pointer network and Asynchronous Advantage Actor-Critic (A3C) is proposed, which utilizes the characteristics of the pointer network in handling variable-length sequence problems to solve it and trains the pointer network with the A3C algorithm for improved performance. Moreover, this paper introduces the joint optimization of multiple metrics, including energy consumption and latency. Experimental comparative analysis demonstrates that the proposed scheme outperforms other schemes and can reduce time and energy consumption.