Filling Action Selection Reinforcement Learning Algorithm for Safer Autonomous Driving in Multi-Traffic Scenes

Learning-based algorithms are gradually emerging in the field of autonomous driving due to their powerful data processing capabilities. Researchers in the field of intelligent vehicle planning and decision-making are gradually using reinforcement learning algorithms to solve related problems. The safety research of reinforcement learning algorithms is significant and widely concerned. The main reason for the safety problem of the existing reinforcement learning algorithm is that there is still a bias in the safety judgment of the current environment, and it is impossible to make directional improvements by modifying the network and training method. In this paper, an action judgment network is designed as a standard to select the optimal action, which can assist the algorithm to judge environmental safety more deeply. Firstly, the action judgment network takes the state space and action as input, and the output is the safety state of the vehicle after the action. Secondly, this work establishes the required database to train the action judgment network through deep learning and achieves the highest accuracy of 98%. Finally, the proposed algorithm is tested in three scenarios: single-lane, intersection, and roundabout. This algorithm can judge the actions according to the reinforcement learning q value table order until the optimal and safe action is selected. The results show that the newly proposed algorithm can greatly improve the safety of the algorithm without affecting vehicle speed.