Analyse des comportements agents par agrégation aux comportements humains

Multi-agent simulations, used in the field of immersive virtual environments, implement agents that reproduce human behaviors and interact with them. In this paper, we propose a generic method based on data analysis to evaluate individual agents behaviors. It uses a simulation logs clustering to extract human behaviors that serve as the reference behaviors to aggregate agents on humans clusters. It is then possible to analyze the capacities, lacks, and errors of the agents behaviors by comparing them to humans ones. We complete this analysis with a subjective evaluation based on a questionnaire filled by human annotators to draw categories of users, making their behaviors explicit. We apply our method to immersive driving simulation.