Design and Implementation of an Adaptive Cruise Control System Based on Supervised Actor-Critic Learning

A novel adaptive cruise control (ACC) system is proposed in this paper. A hierarchical control framework is adopted for the adaptive cruise control problem. For the upper level, a supervised actor-critic (SAC) reinforcement learning approach is presented to obtain the desired acceleration controller. In the lower level, throttle and brake controllers calculate the required throttle and/or brake signals based on the desired longitudinal acceleration. Feed-forward neural networks are used to implement the actor and critic components of the SAC learning algorithm. An online learning mechanism is introduced to implement the SAC training process. dSPACE simulator is used to verify the effectiveness of the ACC system. Typical emergency braking scenario is simulated to test the adaptability of the ACC system. Road condition change (e.g. wintry or wet conditions) simulation is first investigated to evaluate the robustness of the ACC system. Performance of the proposed ACC system is proved to be very practical.