This thesis compares alternative and proposes new candidate algorithms for the online calibration of Dynamic Traffic Assignment (DTA). The thesis presents two formulations to on-line calibration: 1) The classical statespace formulation and 2) The direct optimization formulation. Extended Kalman Filter (EKF) is presented and validated under the state-space formulation. Pattern Search (PS), Conjugate Gradient Method (CG) and Gradient Descent (GD) are presented and validated under the direct optimization formulation. The feasibility of the approach is demonstrated by showing superior accuracy performance over alternative DTA model with limited calibration capabilities. Although numerically promising, the computational complexity of these base-line algorithms remain high and their application to large networks is still questionable. To address the issue of scalability, this thesis proposes novel extensions of the aforementioned GD and EKF algorithms. On the side of algorithmic advancement, the Partitioned Simultaneous Perturbation (PSP) method is proposed to overcome the computational burden associated with the Jacobian approximation within GD and EKF algorithms. PSP-GD and PSP-EKF prove to be capable of producing prediction results that are comparable to that of the GD and EKF, despite achieving speed performance that are orders of magnitude faster. On the side of algorithmic implementation, the computational burden of EKF and GD are distributed onto multiple processors. The feasibility and effectiveness of the Para-GD and Para-EKF algorithms are demonstrated and it is concluded that that distributed computing significantly increases the overall calibration speed.
展开▼
