Uniform TitleCalibration of traffic simulation models using Simultaneous Perturbation Stochastic Approximation (SPSA) method extended through Bayesian sampling methodology
NameLee, Jung-Beom (author), Ozbay, Kaan (chair), Nassif, Hani (internal member), Boile, Maria (internal member), Brail, Richard (outside member), Rutgers University, Graduate School - New Brunswick,
SubjectCivil and Environmental Engineering,
Traffic engineering--Mathematical models,
Traffic flow--Simulation methods
DescriptionThe main goal of this dissertation is to propose a new methodology for the calibration of traffic simulation models. Simulation is useful in representing complex real-world systems, and many alternatives can be compared via different system designs. However, to evaluate road conditions accurately, the selection of model parameters to be calibrated and the calibration methodology are very important aspects of the overall simulation modeling process.
One of the key elements of this dissertation is the application of the Simultaneous Perturbation Stochastic Approximation (SPSA) algorithm (Spall (1992))--one of the well-known stochastic approximation (SA) algorithms, to determine optimal model parameters. The SPSA algorithm has an inherent advantage that can be exploited in both stochastic gradient and gradient-free settings; it can also be applied to solve optimization problems that have a large number of variables.
One of the main distinctions between this study and previous studies is with regards to calibration while considering a wide range of all likely demand conditions. Previous studies on calibration have focused on minimizing a deterministic objective function, which is the sum of the relative error between the observed data and the simulation output from a certain time period in a typical day. Even though this approach can be considered a calibration that uses data obtained at one point in time, this type of calibration approach cannot capture a realistic distribution of all possible traffic conditions. Thus, a more general calibration methodology needs to be implemented--one that enables use with any traffic condition. In this dissertation, we propose the Bayesian sampling approach, in conjunction with the application of the SPSA stochastic optimization method, which enables the modeler to enhance the theoretic application to consider statistical data distribution. Thus, this proposed new and advanced methodology makes it possible to overcome the limitations of previous calibration studies.
Testing the methodology for larger networks, as well as for other microscopic traffic simulation tools such as CORSIM or VISSIM, are future research tasks. In the future, other simulation parameters and more extensive data sets can be used to test the strengths and weaknesses of the proposed calibration methodology.
NoteIncludes bibliographical references (p. 144-148).
CollectionGraduate School - New Brunswick Electronic Theses and Dissertations
Organization NameRutgers, The State University of New Jersey
RightsThe author owns the copyright to this work.