Recherche sur des méthodes d'optimisation pour la mise en place de modèles intégrés de transport et usage des sols

Keywords :

• Tranus
• usage de sol
• calibration
• optimisation
• land use

Land use and transportation integrated (LUTI) models aim at representing the complex interactions between land use and transportation offer and demand within a territory. They are principally used to assess different alternative planning scenarios, via the simulation of their tendential impacts on patterns of land use and travel behaviour. Setting up a LUTI model requires the estimation of several types of parameters to reproduce as closely as possible, observations gathered on the studied area (socio-economic data, transport surveys, etc.). The vast majority of available calibration approaches are semi-automatic and estimate one subset of parameters at a time, without a global integrated estimation.
In this work, we improve the calibration procedure of Tranus, one of the most widely used LUTI models, by developing tools for the automatic and simultaneous estimation of parameters. Among the improvements proposed we replace the inner loop estimation of endogenous parameters (known as shadow prices) by a proper optimisation procedure. To do so, we carefully inspect the mathematics and micro-economic theories involved in the computation of the various model equations. To propose an efficient optimisation solution, we decouple the entire optimisation problem into equivalent smaller problems. The validation of our optimisation algorithm is then performed in synthetic models where the optimal set of parameters is known.
Second, in our goal to develop a fully integrated automatic calibration, we developed an integrated estimation scheme for the shadow prices and a subset of hard to calibrate parameters. The scheme is shown to outperform calibration quality achieved by the classical approach, even when carried out by experts. We also propose a sensitivity analysis to identify influential parameters, this is then coupled with an optimisation algorithm to improve the calibration of the selected parameters.
Third, we challenge the classical viewpoint adopted by Tranus and various other LUTI models, that calibration should lead to model parameters for which the model output perfectly fits observed data. This may indeed cause the risk of producing overfitting (as for Tranus, by using too many shadow price parameters), which will in turn undermine the models' predictive capabilities. We thus propose a model selection scheme that aims at achieving a good compromise between the complexity of the model (in our case, the number of shadow prices) and the goodness of fit of model outputs to observations. Our experiments show that at least two thirds of shadow prices may be dropped from the model while still giving a near perfect fit to observations.

The contribution outlined above are demonstrated on Tranus models and data from three metropolitan areas, in the USA and Europe.

Directors:

• Mr Peter Sturm (INRIA )
• Mr Arthur Vidard (INRIA )

Raporteurs:

• Mr Vincent Hilaire (Université de Technologie de Belfort-Montbéliard )
• Mr Michael Batty (University College London )

Examinators:

• Mr Nicolas Coulombel (Université Paris-Est )
• Mr Carlos Canudas de Wit (CNRS )
• Mr Tomas de la Barra (Universidad Central de Venezuela )