**IRIT – Université Toulouse 1 – Capitole**

** Keywords**

calibration, simulation, multi-agent system, AMAS

In many fields of science and engineering, simulation is a key part in understanding phenomena or predicting their future evolution. It is also a useful tool for planners and decision-makers in order to guide them in their decisions. A simulation is a computer model of a real-world system that contains entities in interaction and that is used to understand and/or predict the evolution of the system it represents.

In order to be as close as possible to real phenomena, simulation models have to be calibrated. Several different calibration methods exist, from classical optimization methods to multi-agent systems and data assimilation. As most simulation models are complex systems, usual calibration methods are not really appropriate to account for dynamics changes. One way to handle those changes is to tune parameters values while the simulation is running, using data observed on the real system. One additional constraint is that models are seen as black boxes, i.e. the calibration system has no insight of the inner workings of the model. This means that the calibration system has to learn the influence of each parameter on each model observation.

This work proposes a new online calibration method, CALICOBA, based on adaptive multi-agent systems, that aims to solve these problems. It features two kinds of agents: parameters and objectives. The role of objective agents is to estimate the distance of observations from the real data. Parameter agents take these distances to evaluate the best value the corresponding model parameter should take in order to minimize the distance computed by objective agents. Parameter agents have to learn their influence on each objective agent in order to compute the best value.

For instance, in the case of a simple traffic model with two parameters (maximum vehicle speed and reaction time) and two observables (traffic density and mean vehicle speed). Lets assume that there is a data source for the actual roads represented by this model. The role of CALICOBA would be to take in the observed data and compare it to each observable in order to compute a distance for each. With this information, the parameter agents have to determine in which direction to modify their values in order to decrease these distances.

As this is still a work in progress, the system has yet to be tested on traffic models, but it is planned for the near future.

**Scientific goals**

- Online self-calibration

- Automatic learning of model input/output interactions

**Contact**

Damien Vergnet – damien.vergnet_at_irit.fr

Frédéric Amblard – frederic.amblard_at_irit.fr

Elsy Kaddoum – elsy.kaddoum_at_irit.fr

Nicolas Verstaevel – nicolas.verstaevel_at_irit.fr