Calibration of mobility and traffic simulation models through machine learning

Urban mobility is one of the main elements of intelligent transportation. In this context, computational mobility and traffic simulation models are harnessed by traffic engineers. Nevertheless, to obtain a scenario similar to real life needs to change parameters in the simulator, this process is iterative and time-consuming. It is traditionally done manually. These simulations help to enhance the traffic flow and predict congestion if something in the network changes. Simulation of Urban Mobility (SUMO) is very popular in the world of traffic simulation. This software is a package of microscopic traffic and open source that simulates urban network behavior. However, these simulations are computationally expensive because of problems in the size of the scenario and the number of vehicles. The number of cars in a system can take a long processing time. This research proposes a methodology to automatically calibrate traffic simulations by counting cars in real time to obtain precise data of input of the simulator. First, many simulations are done to create an extensive dataset of examples by using different volumes of vehicles in entry lanes and probabilities in intersection lanes. Then data is interchanged from input/output to output/input to train the models. It is applied different machine learning techniques, such as Artificial Neural Networks, Random Forest (RF), and k-Nearest Neighbors (kNN) that are capable of estimating simulation results. It is presented with another option for calibration that combines machine learning models and a genetic algorithm if the proposed method does not work well. Ibarra city was selected as the main for calibration and two alternative scenarios with high prevalence in urban areas as well as the fact that their network structures differ from one another. Results have shown Neural Networks have better performance in the first scenario to predict input values to the simulator. In the second scenario, Neural Networks also had better performance, however, the results were not so accurate. That is why the alternative that combines machine learning models with a genetic algorithm was performed. kNN achieved better performance in predicting the outputs from the simulator without its execution. Once a model with high precision was developed, a genetic algorithm was implemented to obtain the input values of the simulation having the counting of cars in intersections.

• Autor principal: De la Cruz Paucar, Franklin Steven (author)
• Formato: bachelorThesis
• Idioma: eng
• Publicado em: 2023
• Assuntos: Redes neuronales artificiales; Aprendizaje automático; Algoritmo genético; Artificial neural networks; Machine learning; Genetic algorithm
• Acesso em linha: http://repositorio.yachaytech.edu.ec/handle/123456789/675