Abstract
Driving in dense traffic with human and autonomous drivers is a challengingtask that requires high-level planning and reasoning. Human drivers can achievethis task comfortably, and there has been many efforts to model human driverstrategies. These strategies can be used as inspirations for developingautonomous driving algorithms or to create high-fidelity simulators.Reinforcement learning is a common tool to model driver policies, butconventional training of these models can be computationally expensive andtime-consuming. To address this issue, in this paper, we propose ``skill-based"hierarchical driving strategies, where motion primitives, i.e. skills, aredesigned and used as high-level actions. This reduces the training time forapplications that require multiple models with varying behavior. Simulationresults in a merging scenario demonstrate that the proposed approach yieldsdriver models that achieve higher performance with less training compared tobaseline reinforcement learning methods.