Stochastic Driver Modeling and Validation with Traffic Data

This paper describes a stochastic modeling approach for predicting driver responses in highway traffic. Different from existing approaches in the literature, the proposed modeling framework allows simultaneous decision making for multiple drivers (>100), in a computationally feasible manner, instead of modeling the decisions of an ego driver and assuming a predetermined driving pattern for other drivers in a given scenario. This is achieved by a unique combination of hierarchical game theory, which is used to model strategic decision making, and stochastic reinforcement learning, which is employed to model multi-move decision making. The proposed approach can be utilized to create high fidelity traffic simulators, which can be used to facilitate the validation of autonomous driving control algorithms by providing a safe and relatively fast environment for initial assessment and tuning. What makes the proposed approach appealing especially for autonomous driving research is that the driver models are strategic, meaning that their responses are based on predicted actions of other intelligent agents in the traffic scenario, where these agents can be human drivers or autonomous vehicles. Therefore, these models can be used to create traffic models with multiple human-machine interactions. To evaluate the fidelity of the framework, created stochastic driver models are compared with real driving patterns, processed from the traffic data collected by US Federal Highway Administration on US101 (Hollywood Freeway) on June 15th, 2005.