Driving simulators are used by car manufacturers to develop driver aid systems, and to carry out experiments on ergonomics, design, or drivers' behavior : the objective is to improve the vehicles safety, and to reduce delays and development costs. In simulators, the driver is immersed in a simulated traffic, which realism is crucial for the validity of the results : the more realistic the environment is perceived, the more significant the drivers' reactions are. In individual-centered approaches, like traffic simulation in our case, the agents' behavioral variety is an important realism criteria. The behaviors also have to be consistent : if aberrant situations appear, the simulation realism is strongly impacted. In this work, we adressed the issue of the simultaneous influence of these two elements. Furthermore, this work was led in an industrial context : it took place at Renault, in collaboration with the Computer Science Laboratory of Lille. This involved taking into account specific needs : the designed tools had to allow experts specifying various and consistent behaviors, and final users easily using them. In this work, we propose a behavioral differentiation model, which provides the basis for a generic and non-intrusive tool allowing an out-of-the-agent design. The model involves three dimensions. First, it describes the agents' behaviors using norms. They provide a behavioral pattern during conception, and a compliance reference during execution. Then, the generation process of the behaviors allows the creation of deviant or violating agents, by influencing the determinism of the mechanism. Finally, the norms can be inferred from previous simulations records or real data, providing an analysis tool of the results and allowing automating the model configuration. We applied the model to the traffic simulation in scaner, the driving simulation software developed and used at the Technical Center for Simulation of Renault. The developments carried out during the thesis introduced driving styles in the traffic (e.g. cautious or aggressive drivers) specified using norms. The use of norms allows populating the environment easily and in an automated way. These developments are already included in the commercial version of the software. The behavioral realism of the traffic was improved, and the experimentations show how the model contributes to the variety and the representativeness of the produced behaviors.
