Achieving Realism in Traffic Simulations: Performance of a Cognitive Behavior Model on the Waymo Open Sim Agent Challenge

Modeling human behavior in traffic is crucial for the development and testing of autonomous vehicles. Traditional rule-based models, while effective in ensuring collision avoidance and adherence to traffic rules, fail to capture the full spectrum of human behavior. Conversely, data-driven models utilizing machine learning techniques, such as neural networks, offer more realistic simulations but require vast amounts of data, especially for rare but critical situations like accidents. This paper introduces cogniBOT, a cognitive behavior model that integrates both rule-based and data-driven approaches by incorporating existing knowledge about human information processing. Our model, which uses significantly fewer parameters than purely data-driven models, is capable of adapting to both common and rare traffic situations. We demonstrate that cogniBOT, without being fitting on the provided training data set, achieves competitive performance in the Waymo Open Sim Agent Challenge, coming close to state-of-the-art data-driven models in terms of realism and interaction metrics. In contrast to its competitors, our general-purpose road user model has a broad range of applications while maintaining transparency and explainability.