Learning the human longitudinal control behavior with a modular hierarchical Bayesian Mixture-of-Behaviors model

Modeling drivers' behavior is believed to be essential for the rapid prototyping of error-compensating assistance systems. Various authors proposed control-theoretic and production-system models. These models are handcrafted in a top-down software engineering process. Here we propose a machine-learning alternative by estimating stochastic driver models from behavior traces. They are more robust than their non-stochastic predecessors. In this paper we present a Bayesian Autonomous Driver Mixture-of-Behaviors (BAD MoB) model for the longitudinal control of human drivers in an inner-city traffic scenario. It is learnt on the basis of multivariate time-series obtained in simulator studies. Percepts relevant for longitudinal control were included in the model by a structure-learning method using Bayesian information criteria. Besides mimicking human driver behavior we suggest using the model for prototyping intelligent assistance systems with human-like behavior.