Ant Colony Systems and the Calibration of Multi-Agent Simulations: a New Approach

International audienceIn this paper, we propose a new approach for the exploration of the parameter space of agent-based models: Adaptative Dichotomic Optimization. Agent-based models are generally characterized by a great number of parameters, a lot of which cannot be evaluated with the current knowledge about the real system. The aim of the work is to provide tools for the calibration of these models, which consists in nding the optimal set of parameters for a given criterion. The criterion can be for example that the model achieves a specic function optimally or that the results of the simulation are as close as possible to experimental data. Our approach is based on the partition of the parameter space (the interval of variation of each variable is divided into a nite number of intervals) and on a parallel exploration of the various parameters by the agents of the model. The navigation in the parameter space is done bygrouping or dividing adaptively some of the intervals, according to an algorithm which is adapted from Ant Colony Systems