TRAVESTI stands for “TRAffic Volume Estimation by Spatio-Temporal Inference.”

ANR SYSCOMM Project N°ANR-08-SYSC-017

This project addresses the problem of modelling large scale complex systems to provide predictions of their macroscopic behaviour. For application purpose, we focus here on the particular problem of the real-time prediction of traffic conditions on a road network.

Car traffic is a typical complex system which exhibits emerging phenomena such as jam formation and long distance interactions throughout a network. In particular we focus on the analysis of the traffic patterns delivered by the METROPOLIS traffic simulation software, in order to set up a prediction method based on a statistical physics modelling and message-passing algorithms.

We plan to study two types of traffic conditions:

• realistic conditions, based either on synthetic data provided by computer simulations or on historically recorded field data from fixed or moving sensors as in a real-world environment,

• futurist traffic conditions in the perspective of the automated road.

As usual for complex systems, the issue is to extract from local interaction rules, a macroscopic representation of the behaviour of the system. By definition of a complex system, it exhibits macroscopic phenomena that cannot be directly deduced from individual microscopic behaviours and the understanding of the overall behaviour of the system resides in the identification of relevant variables and macroscopic structures.

In the target application, we consider two complementary points of view:

• In the first one, developed in METROPOLIS, individual agents are users of the network, each one trying to minimize a utility function, typically generalized travel costs. The competition for limited resources (the maximal capacity of the road network) is at the origin of the interaction between agents. As a result, traffic congestion patterns emerge and the inverse problem at hand is to find the correct parameters to produce realistic situations.

• In the second one, the elementary units are the road sections, considered at successive discrete times, and one wants to take advantage of statistical correlations between the states of these sections. The idea is to produce a macroscopic description of the network in order to be able to make traffic predictions.

Our proposal focuses on developing a new approach to this second point of view with large scale and real time constraints. The tools which will be used are

• theoretical tools from probability and statistical physics.

• algorithmic tools, for continuous optimization and message-passing algorithms which have been flourishing lately.

• statistical tools for data analyses and pattern recognition.

To produce the data, METROPOLIS will be our basic tool, serving either as a proxy for real world data or for creating synthetic conditions.