Automatic calibration of traffic microsimulations with artificial neural networks.

• Ano de defesa: 2019
• Autor(a) principal: Daguano, Rodrigo França
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Artificial neural network; Calibration automation; Redes neurais; Regression model; Simulação de sistemas; Sistemas inteligentes de transporte; Smart transportation; Traffic simulation; Vissim
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/3/3142/tde-27052022-101649/

In the topic of Smart Transportation, urban planning took great advantage from computational simulation tools for traffic. In microsimulations, one important step is calibration, which is accomplished by tuning the values of simulation inputs, in order to match its internal metrics with those from real-world data. The process is iterative, time-consuming and is traditionally done manually by a traffic engineer. This research proposes a methodology to automatically calibrate traffic simulations. Initially a large number of simulations are run to create an extensive dataset of examples. Then, the dataset is used for training Artificial Neural Networks that are capable of estimating the simulation inputs that deliver the target output metrics, thus calibrating the simulations upon request of specific scenarios. Validation experiments were conducted to calibrate the routing and flow setup of simulations, and in these experiments it has been verified a high correlation, above 80%, between the estimates from the Neural Networks and the desired values for the input variables of the simulator, therefore validating the proposed methodology and the capabilities of automation and scalability of the calibration process. The evaluation of the Neural Networks also delivers the metrics for each individual input variable, such as vehicle volumes and route decisions, thus allowing the user to choose or ignore the estimates for those variables with poor performance and instead proceed to manual calibration. Finally, two experiments investigated the calibration capabilities of driving behavior parameters, a more abstract type of variable. For this type, results were observed with acceptable accuracy for about half of the parameters. In this case, it is the users choice to ignore the variables with the worst performances and use those with acceptable performances as a starting point for refinement. The proposed methodology has been shown to be capable of estimating with sufficient accuracy a significant part of the calibration parameters, thus reducing the workload of a traffic engineer and allowing more dedication to the work of scenario analysis.