Parameter estimation in urban models: Theory and application to a land use transport interaction model of the Sacramento, California region

Large scale urban models are desirable for transportation planning and policy making because they simulate and thereby help understand and take into account the interactions between the transportation system and the spatial economic system. These models draw on theories from various disciplines to understand the nature of the entire urban system and how it might evolve. To manage this complexity, the models can be divided into sub-models. Thus parts of the model can be considered individually as can specific interactions between two or more parts.; It is common to use parameter values in an urban model from other models that are somewhat different than the modelling system used for forecasting. Sometimes these other models are different mathematically, but other times they are only the sub-models of the larger model, and the difference is that their parameters are estimated from observed data, where as in the final modelling system the circular interconnections mean that most data is synthesized by other sub-models.; These theoretical differences between models considered individually and models combined into a system strongly suggest that it is appropriate to ensure that the entire model system is accurate. This can be done through an overall calibration of the entire urban model.; These issues are investigated in the context of a comprehensive land use transportation interaction model of the Sacramento, California region based on the MEPLAN mathematical framework. The model was constructed from its sub-models and a final calibration of the entire system was performed by hand, to provide a working modelling system and to gain knowledge about the characteristics of the modelling system. The model was used to analyse transportation policy in the Sacramento region to gain an understanding of the strengths and weaknesses of the model and its applicability for policy analysis.; The knowledge gained in hand-calibrating and using the model guided the development and use of an automatic parameter estimation procedure (and software), that runs the model many times searching for the “best” parameters given certain assumptions. The automated search process uses a customized version of Newton's method, and hence requires a linearization of the model using numerical derivatives. The automated search process was able to improve the goodness-of-fit measure.; The weights in the goodness-of-fit measure were subject to ad-hoc adjustment based on an understanding of the purpose and use of the model. The sensitivity of the parameters to small changes in the weights can be calculated based on the numerical linearization of the model and on the convergence criteria. This sensitivity information provides guidance towards selecting weights, but it is also valuable in suggesting the causes of lack-of-fit and hence suggesting changes in the design of the modelling system. The estimation software was extended to allow this information to be explored interactively, by showing the largest changes in both targets and parameters that would result from a change in one target's weight.