Individualized SAC Car-Following Strategies Considering the Characteristics of the Driver

Increasing the degree of individuality of the autopilot and adapting it to the habits of drivers with different driving styles will help to increase occupant acceptance of the autopilot function. Inspired by the Twin Delayed Deep Deterministic policy gradient algorithm(TD3) algorithm to increase action spontaneity, this paper proposes a Soft Actor-Critic(SAC) based personalized following control strategy to increase the degree of strategy personalization through driver data. In order to obtain real driver data, this paper collected driving data based on driver-in-the-loop experiments conducted on a simulated driving platform, and selected data from three drivers with distinctive driving characteristics for model training. A continuous action space model was developed by vehicle following kinematics. A temporal Gate Recurrent Unit (GRU) based reference model is trained to receive temporal state signals and output acceleration actions according to the current state. In this paper, we introduce temporal imitation learning into the SAC algorithm by weighting the average of the output actions of the reference model and the output of the SAC strategy network to improve the personalization of the decision algorithm. The reward function has been designed to take into account the safety, comfort and pleasant nature of the following process. Simulation results based on the CARLA simulator show that the personalised following control strategy proposed in this paper is able to learn different driver characteristics in terms of overall style, while ensuring the stability and safety of the vehicle autonomous following process.