A systematic approach to the calibration of traffic assignment models

The urban transportation model predicts flows on a transportation network as a function of the urban system containing the network and the characteristics of the transportation network. It comprises four steps: (1) trip generation, (2) trip distribution, (3) modal split, and (4) traffic assignment. The fourth step, traffic assignment, predicts routes (paths) used between each origin-destination pair on the transportation network.; Calibration of traffic assignment models is acknowledged as a perquisite to their application. Such calibration efforts have traditionally been ad hoc. They have fallen into one of five categories: (1) modification of the network representation, (2) adjustment of travel demand, (3) selecting the traffic assignment method and assumptions, (4) adjustment of the traffic dispersion parameter for stochastic assignments, and (5) estimating congestion function parameters.; Some efforts have recently been made towards systematizing the calibration of traffic assignment. These efforts can be categorized as mathematical programming and heuristic procedures. Mathematical programming has been traditionally applied in the estimation of congestion function parameters. If observed link flows and travel times are available, least squares are applied.; In practice, however, it is hard to measure link travel times. Proposed is a generic method I term network loading (NL). This method simultaneously estimates the link flows and congestion function parameters. If equilibrium link flows are used in parameter estimation and prediction, we have what I call the user-equilibrium (UE) model. A particular implementation of the UE model is the Entropy-Maximizing (EM) model. The bilevel formulation by Suh et al. (1990) is a similar approach that I cover in this research. It has an upper-level objective function that minimizes the error in estimated link flows. The lower-level problem is the formulation of the UE assignment.; Despite their elaborate formulations, the mathematical programming methods are difficult to implement. The difficulty increases with the size of the network. Human-based expertise and heuristics provide a viable alternative approach. Fricker (1989) proposes the parameter adjustment (PA) method for the calibration of congestion function parameters. He also proposes the direct impedance adjustment (DIA) and simultaneous link speed adjustment (SLSA) methods for modifying link free-flow travel times and speeds. The three proposed procedures require only observed link flows.; The three heuristic procedures are tested on two networks, the fictitious Archerville and the real Eindhoven networks. The experiments on the Archerville network are controlled for extraneous sources of error while those on the Eindhoven network are not. The Eindhoven network is supplemented with synthetic data and more controlled tests are conducted.; The conclusions of the experiments are as follows: (1) PA is very sensitive to parameter starting values and level of congestion on the network. It performs well only on highly detailed networks. (2) DIA and SLSA improve traffic assignment performance in terms of replicating actual link flows. DIA, however, does not always perform very well. (3) DIA and SLSA perform better at higher levels of network detail. (4) DIA and SLSA can be applied to all or parts of a network. (5) SLSA is clearly superior to DIA. Accordingly, it recommended for use by transportation planners.