Enhancing vehicle-to-vehicle communication in urban scenarios through artificial intelligence

In this work, two contributions are presented to improve vehicular communications through artificial intelligence models. The first contribution analyzes the impact of synthetic data generation techniques on a dataset built from routing metrics, combining OpenStreetMap, SUMO, OMNeT++, and Veins. Since the dataset was imbalanced, balancing and augmentation techniques using SMOTE and GAN were explored to obtain a more robust model. In addition, advanced algorithms were employed for hyperparameter optimization and relevant metric selection. The creation of the dataset required 840 manual hours and 3,600 simulation hours. Data augmentation made it possible to increase the amount of information without exhaustive simulations. The results show that SMOTE achieved an AUC of 0.925, outperforming the 0.887 reported in previous studies, thereby improving model performance. The second contribution analyzes the impact of different congestion control mechanisms in vehicular networks through simulations on realistic maps of Houston, Liverpool, and Rio de Janeiro. These cities were selected according to their low, medium, and high urban complexity to ensure the model’s generalization capability. To evaluate the efficiency of each mechanism, key metrics such as packet delivery ratio (PDR), delay, and throughput were analyzed. Three schemes were evaluated: a scenario without congestion control, an EDCA-based approach, and a machine learning–based mechanism using CatBoost. The results demonstrated that the machine learning model outperforms both EDCA and the no-control scenario. In Rio de Janeiro, a PDR of 60–66% was achieved compared to 35–40% with EDCA. In Houston, the delay was below 0.015 seconds compared to 0.08 seconds without control. Throughput reached 2.9 kbps versus 2.2 kbps with EDCA, with an inference time of 4.9 microseconds, achieving significant improvements across all evaluated metrics.

• Tác giả chính: Carlosama Perachimba, Luis David (author)
• Định dạng: bachelorThesis
• Được phát hành: 2026
• Những chủ đề: VANETs; Ad Hoc; Machine learning
• Truy cập trực tuyến: https://repositorio.yachaytech.edu.ec/handle/123456789/1049