Automated Generation of Various and Consistent Populations in Multi-Agent Simulations

The variety and consistency of the agents behaviors greatly influences the realism in multi-agent simulations, and designing scenarios that simultaneously take into account both aspects is a complex task. To address this issue, we propose an approach to automatically create populations using sample data. It facilitates the designers tasks, and variety as well as consistency issues are handled by the generation model. The proposed approach is based on a behavioral differentiation model that describes the behaviors of agents using norms. To automatically configure this model, we propose an inference mechanism based on Kohonen networks and estimation distribution functions. We then introduce agents generators that can create a specified population, and are automatically configured by the inferred norms. The approach has been evaluated in traffic simulation, in association with a commercial software. Experimental results show that it allows to accurately reproduce the populations represented in sample data.