Modeling Driver Lane Change Behavior Using Inverse Reinforcement Learning

This paper presents an approach for modeling driver lane change behavior using inverse reinforcement learning (IRL). The driver behaviors are assumed to be the solution of an optimal control problem with unknown weights that simultaneously optimizes several objectives and follows a Bezier curve. The IRL is then used to identify the weights that best explain the driver's preference. To test the efficacy of the proposed methodology, expert trajectories are generated using simulation, which serves as the data set for this model training. Test results demonstrate the capability of the proposed methodology to reproduce each driver's individual trajectory, which can then be fed into low level motion control to provide personalized autonomous driving.