|M.Sc Student||Tatiana Kolechkina|
|Subject||Off-Line Estimation of Dynamic Origin-Destination Trip|
|Department||Department of Civil and Environmental Engineering||Supervisor||Professor Toledo Tomer|
|Full Thesis text|
The origin-destination (OD) demand matrix is a key input required for Dynamic Traffic Assignment (DTA) models utilized for evaluation of various technologies and strategies for intelligent traffic network management. Since gathering OD demand information directly by conducting surveys is very costly and time consuming, most commonly OD matrices are estimated indirectly using available traffic measurements. The OD estimation process efficiency results in more realistic behavior of the DTA model and consequently in more accurate traffic analysis.
The aim of this research is to develop an efficient solution methodology for the problem of off-line estimation of dynamic OD matrices using historical OD information and link traffic counts. The given OD estimation problem is formulated as a single-level optimization problem. Special interest is given to the assignment map, which describes how the travel demand is transferred onto the network links. Because of congestion, the assignment proportions depend on the unknown travel demand. To overcome this problem of circular interdependency between the OD demand and the assignment proportions, we propose to use a linear approximation of the assignment matrix.
The developed solution strategy is applied in several iterative algorithms based on different optimization techniques: Gradient and quasi-Newton methods and simultaneous perturbation stochastic approximation (SPSA) technique. In addition, this work presents a modification of the heuristic proposed by Lundgren et al. (2006). For the traffic assignment modeling part in the algorithms’ implementation, the mesoscopic traffic simulation tool Mezzo has been used.
Practical tests on a prototypical network verify the correctness and effectiveness of the developed solution approach under a variety of demand and congestion settings. Demonstration on the real network model confirms the applicability of the proposed methodology to large and complex networks and its feasibility to replicate “true” demand values with sufficient accuracy.