CNRS et Laboratoire d'Economie des Transports, Lyon France
November 22, 2007, 11:15, Room GC B3 424 (click here for the map)
In this talk we present a methodology to calibrate a class of dynamic traffic simulation models with real data acquired from traffic counts and travel time measurements acquired from GPS devices. Calibration can be described as an optimization process with an objective function that depends on simulation outputs. That function is often noisy because simulations include stochasticity and costly to compute because a single evaluation implies a complete microscopic simulation, a task which can range from a few minutes to up to an hour of computer time. We explore the use of population based optimization algorithms such as genetic algorithms in general and Covariance Matrix Adaptation-Evolutionary Strategy (CMA-ES) in particular. We propose BBOG, a framework designed to swap different algorithms in order to compare their performances for traffic calibration. The framework is tested using CMA-ES to calibrate METROPOLIS simulations running on a cluster of Linux servers. As a proof of concept, we present the results of the calibration procedure with an application to the toy network of Sioux Falls. Two important findings are derived from this exercise. Firstly, state of the art evolutionary algorithms are promising search methods to calibrate traffic simulations, at least when the number of parameters is relatively small. Secondly, a unified framework to compare algorithms provides a richness of analysis that is usually missing from the traditional application of a given algorithm to a particular traffic problem.
Fabrice Marchal est charge de recherche CNRS au Laboratoire d'Economie des Transports, Lyon. Ses recherches portent principalement sur la simulation des systemes de transports a grande echelle (modeles METROPOLIS et MATSIM) ainsi que sur les modeles integres transport-occupation du sol.