Task-Offloading Strategy Based on Performance Prediction in Vehicular Edge Computing

Abstract

In vehicular edge computing, network performance and computing resources dynamically change, and vehicles should find the optimal strategy for offloading their tasks to servers to achieve a rapid computing service. In this paper, we address the multi-layered vehicle edge-computing framework, where each vehicle can choose one of three strategies for task offloading. For the best offloading performance, we propose a prediction-based task-offloading scheme for the vehicles, in which a deep-learning model is designed to predict the task-offloading result (success/failure) and service delay, and then the predicted strategy with successful task offloading and minimum service delay is chosen as the final offloading strategy. In the proposed model, an automatic feature-generation model based on CNN is proposed to capture the intersection of features to generate new features, avoiding the performance instability caused by manually designed features. The simulation results demonstrate that each part of the proposed model has an important impact on the prediction accuracy, and the proposed scheme has the higher Area Under Curve (AUC) than other methods. Compared with SVM- and MLP-based methods, the proposed scheme has the average failure rate decreased by 21.2% and 6.3%, respectively. It can be seen that our prediction-based scheme can effectively deal with dynamic changes in network performance and computing resources.

Keywords

• vehicular edge computing
• task offloading
• performance prediction
• deep learning
• service delay