Research Information System
2nd International Conference on Optimization and Data Science in Industrial Engineering, ODSIE 2024, Virtual, Online, 7 - 08 November 2024, vol.2482 CCIS, pp.217-231, (Full Text)
This study investigates strategies for optimizing traffic speed prediction and adaptive traffic control to address urban congestion, focusing on improving training efficiency without sacrificing accuracy. Recent research has leveraged deep learning methods, particularly Graph Neural Networks (GNNs), but these approaches often face challenges like prolonged training times and high storage demands, especially with large-scale datasets. Using publicly available datasets such as METR-LA and traffic data from TomTom France, this paper explores the impact of adjacency matrix preconditioning on training efficiency and model performance through sensitivity analysis. The findings emphasize the importance of selecting optimal initial values, which are highly sensitive to the dataset used. The proposed methodology for French traffic data demonstrates significant improvements in reducing training time while maintaining predictive accuracy, outperforming traditional methods. These results contribute to enhancing the efficiency of deep learning models for traffic prediction and provide practical recommendations for better urban flow management systems, showcasing how real-time data integration, such as from TomTom, can help mitigate city traffic congestion.