Modelos preditivos comportamentais para otimização de despacho de táxis

• Ano de defesa: 2019
• Autor(a) principal: Otsuka, Breno José Bueno
• Orientador(a): Vivaldini, Kelen Cristiane Teixeira
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de São Carlos; Câmpus São Carlos
• Programa de Pós-Graduação: Programa de Pós-Graduação em Ciência da Computação - PPGCC
• Palavras-chave em Português: Sistema de transporte inteligente; Problema de atribuição táxi-passageiro; Despacho de táxi; Aprendizagem de máquina; Otimização combinatória
• Palavras-chave em Inglês: Intelligent transportation system; Taxi-passenger matching problem; Taxi dispatching; Machine learning; Combinatorial optimization
• Área do conhecimento CNPq: CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::METODOLOGIA E TECNICAS DA COMPUTACAO
• Link de acesso: https://repositorio.ufscar.br/handle/20.500.14289/11511

The popularization of smartphones has given rise to several online taxi-booking applications as a more efficient way to call for a taxi. These applications mediate communication between passengers and taxi drivers, reducing the waiting time of passengers and increasing the reliability of the service. This intermediation consists of the Taxi-Passenger Matching Problem, whose goal is to select the best taxi driver for each passenger, and to solve this problem a taxi dispatch method is used. In this context, the challenge arises to adapt this method to the interests and needs of users. Thus, in this work, a multi-passenger approach was proposed that used users' prediction of behavior with the goal of maximizing the rate of successful assignments. For this, two predictive models were trained using Supervised Machine Learning methods (Logistic Regression and Gradient Boosted Decision Trees), one to estimate the probability of a taxi driver accepting the offer of a request and another to estimate the probability of a request being answered by a taxi driver. In addition, two objective functions, one linear and one non-linear, were implemented to evaluate offer distributions. A heuristic was used for each of the functions, and for the linear it was guaranteed the optimal solution and for the non-linear one not. We also implemented two selection criteria, one based on cancellation of requests by passengers and another by waiting time of users. In numerical simulations with data from the capital of São Paulo made available by Easy Taxi, the proposed taxi dispatching method was tested. The results indicated that the predictive model of acceptance was superior to that of attendance with the objective function non-linear and inferior to linear, in addition it was verified that the selection criterion based on cancellation was slightly superior to the one based on waiting time. The best attendance rate was 74.76% with the predictive model of acceptance with the nonlinear function and the criterion of selection by cancellation. Finally, it was tested the repetition of requests not offered or not accepted, obtaining 76.59% of attendance.