Proposta de aplicação de sistema de inferência neuro-fuzzy para otimização de tráfego

• Ano de defesa: 2005
• Autor(a) principal: Gobbo, Alexandre Fadel
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Centro Federal de Educação Tecnológica do Paraná; Curitiba; Programa de Pós-Graduação em Engenharia Elétrica e Informática Industrial
• Palavras-chave em Português: Engenharia de tráfego; Traffic engineering
• Link de acesso: http://repositorio.utfpr.edu.br/jspui/handle/1/94

It is proposed in this dissertation a traffic optimization model based on neuro-fuzzy inference systems (ANFIS) to be applied in real-time optimization, targeting the architecture of the traffic control system deployed in the city of Curitiba, considering its properties and restrictions. The model presented herein has a similar approach to the best known dynamic traffic optimization system, SCOOT, implementing hill-climbing optimization on a performance index defined as a weighted sum of the links' measures of effectiveness. However, the nature of the traffic flow sampling in Curitiba, by not distinguishing the flow tuning rates, implies the simulation model should be closer to those implemented by offline traffic simulation tools. Transyt, which is one of the most popular offline traffic simulators, was used in order to generate training data and as a reference to validate the results. Real-time optimization based on the mathematical model of Transyt cannot be applied to systems with a large number of intersections in the current state of technology, due to the high computational costs of the algorithms, unless by imposing restrictions on the search space. ANFIS was used in order to capture the knowledge of the simulator, which means to approximate Transyt's outputs. ANFIS was chosen due to its precision and low execution time. Its estimations made it possible the real-time execution of optimization algorithms for a very large number of intersections. Regarding the optimization processes, well-known methods such as EQUISAT and hill-climbing were used. A hybrid optimization method was also validated using a genetic algorithm to provide an initial state for the hill-climbing method.