Driving-style-augmented Aggressiveness-aware Decision Making for Autonomous Vehicles

To achieve safer and more efficient lane-change for autonomous vehicles, this study proposed a novel driving aggressiveness-aware decision-making method based on deep reinforcement learning with considering driving style. Firstly, the driving aggressiveness conception is introduced and formulated based on the analogy with mechanical theory to accurately cognize the driving situation. Meanwhile, the driving aggressiveness distributions of different drivers are analyzed through the cluster of naturalistic driving data, which was demonstrated to be distinguished from the common risk indicator. Subsequently, the driving-style-augmented aggressiveness model is proposed and calibrated by utilizing the classified driving data. Secondly, the driving aggressiveness is employed as the observation states in the deep double Q-network algorithm and is instrumental in designing the reward function. Simultaneously, the trajectory planning module using a quintic polynomial is embedded into reinforcement learning to guarantee the feasibility of lane-change decisions. Finally, the comparative study with the common deep double Q-network and rule-based method is conducted to verify the effectiveness of the proposed framework. The training and test outcomes at the SUMO platform reveal that the proposed model has superior performance in achieving both training efficiency and driving safety, and it can adapt to dynamic driving environments.