Probabilistic Models of Driver Behavior

The BATMobile [1] is one approach to the implementation of a system capable of driving autonomously in normal highway tra c. As with human drivers, the BATMobile would be a superior driver if it were able to make predictions about the actions of other drivers, e ectively giving it the ability to respond before such actions even occur. The ability to predict driver behavior would also be helpful in the creation of accurate models of tra c behavior used in freeway design and analysis. In this report, we discuss the creation of dynamic probabilistic networks (DPNs) that constitute models of driver behavior. Speci cally, we discuss three model structures we have created, the data used to learn the models, and how well the models predict driver behavior. We compare the three models to observe the bene ts of modeling hidden state and using deterministic variables. We also analyze the models to see what aspects of driver behavior the models have learned.