Data-driven probabilistic modeling and verification of human driver behavior

We address the problem of formally verifying quantitative properties of driver models. We first propose a novel stochastic model of the driver behavior based on Convex Markov Chains, i.e., Markov chains in which the transition probabilities are only known to lie in convex uncertainty sets. This formalism captures the intrinsic uncertainty in estimating transition probabilities starting from experimentally-collected data. We then formally verify properties of the model expressed in probabilistic computation tree logic (PCTL). Results show that our approach can correctly predict quantitative information about driver behavior depending on her state, e.g., whether he or she is attentive or distracted.