Virtual worlds play an important role in virtual training scenarios, simulations and games. For users or players interacting with these environments, the worlds should feel immersive, plausible, and often entertaining. Virtual agents, which inhabit these environments, can greatly enhance these effects by conveying the impression of a living world, even if they are implemented in a simple way. In some cases, however, simple reactive behaviour is not sufficient to provide a plausible, meaningful experience. Instead, agents need to perceive, understand, and reason about their current situation to a certain degree before selecting an action. Cognitive agents perform these processes to portray more human-like behaviour for increased realism. However, traditional cognitive architectures require much processing power and are not suitable for interactive applications.

In this talk, the results of PhD work regarding a multi-agent traffic simulation for virtual environments are outlined. The results include a constrained cognitive agent architecture, a multi-layered simulation approach, a semantic road network model, and a virtual perception framework. Agent behaviour is personalized by applying a psychological personality profile, which ensures that agents select different options in a given scenario. At the same time, it guarantees that agents will choose actions that are consistent with their previously observed behaviour. Furthermore, an overview of current research topics regarding agent perception and selective attention is presented alongside relevant background information.