V2X-Based Decentralized Singular Value Decomposition in Dynamic Vehicular Environment

Abstract

The proliferation of road vehicles has led to increased congestion and incidents, necessitating advanced solutions like Intelligent Transport Systems that leverage Vehicle-to-Everything (V2X) communication for enhanced safety and traffic management. However, many current solutions that utilize V2X focus on short-term interaction, such as at an urban intersection. Meanwhile, dynamic traffic challenges the performance of long-term interaction among vehicles, such as machine learning model training. Using an important type of algorithm, singular value decomposition (SVD), as an example, we propose the DGradSVD algorithm, a decentralized SVD method to address the inherent challenges of dynamic, non-centrally controllable vehicular networks. In the evaluation, we investigate the dynamic properties of vehicular networks and the performance of DGradSVD using simulations in SUMO-based synthetic and real-world traffic scenarios. The results highlight the limitations of collaborative algorithms caused by dynamism in traffic, and the proposed algorithm can effectively adapt to such limitations while maintaining model accuracy.