Calibration of Multi-Agent Simulations through a Participatory Experiment

In the context of agent-based simulation, a major issue is to define relevant parameters of the agent model and calibrate them. In this paper, we propose to log and analyse agents behaviours to evaluate their similarity to humans behaviours in an immersive virtual environment. The behaviour archetypes are studied in terms of cluster members in order to identify agent missing behaviours, capacities and errors. This study enables to (1) dismiss invalid parameter sets, (2) calibrate valid simulations and (3) explain lacks in the agent models for further improvement.