Optimal Driving Policies for Autonomous Vehicles Based on Stochastic Drift Counteraction

This paper focuses on high-level control and decision-making for an autonomous car. We develop a hybrid probabilistic model that describes the motion of a car and its surrounding traffic on a two-lane highway/road, where the acceleration of the car and the lane changes serve as control variables. Using approximate dynamic programming (ADP) techniques and an enhanced version of the value iteration algorithm, a control policy is obtained that maximizes the expected time that the car maintains a prescribed minimum (safe) headway. Simulation results for different settings are provided to validate the approach.