Improving the Accuracy and Efficiency of Demand Calibration for Dynamic Traffic Models

Dynamic traffic assignment models involve hundreds of unknown parameters and consume considerable time for computation. Besides, the calibration process, which tries to calibrate demand (and supply) parameters to match the simulation output with real observations, is both methodologically and computationally challenging. Considering availability of the input data and field observations, the core of the OD (Origin-Destination) estimation is formulating it as an optimization problem, with the specification of the objective function and selection of a proper optimization method that can provide accurate and realistic result given problem-related limitations.

While the solution algorithms for simulation-based dynamic OD estimation in the literature are appealing, they all have shortcoming and weaknesses besides their strong points. In this research, we investigate dynamic OD estimation within two aspects; optimal problem/algorithm formulation and problem size reduction.

Former research direction tries to improve the efficiency and accuracy of OD estimation systematically by coupling state-of-the-art solution algorithms into hybrid combinations. We reformulate the OD estimation problem with a vector-valued objective, then dissect existing methods into their fundamental algorithmic design choices and combine them to exploit maximally their benefits. We do this based on an in-depth analysis of the convergence of existing methods through the time and spatial domains.

In the latter approach, the size problem is reduced by defining functions that use socio-demographic data to generate OD matrices. In this method, the calibration is based on the estimation of the (relatively few) parameters that form aggregated demand relations rather than of the (high number of) individual cells of the origin-destination (OD) matrix. In addition to dimension reduction, we examine whether this approach helps avoiding local optimal and maintains better structure of the resulting OD-matrix.