Decision-Making for Autonomous Vehicles in Random Task Scenarios at Unsignalized Intersection Using Deep Reinforcement Learning

This study constructs a decision-making control framework for autonomous ego vehicles (AEVs) based on Soft Actor-Critic (SAC) in a random driving task scenario at an unsignalized intersection. The environment vehicles include both AEV and surrounding vehicles, and the three driving tasks through unsignalized intersections are going straight, left turn, and right turn. Since the driving tasks of AEV and surrounding vehicles are random, the environment is characterized by high uncertainty and difficulty. There are three innovative points in this paper. First, this paper proposes a new Mix-Attention Network based on the attention mechanism. Second, this paper improves the state by introducing a new input quantity to represent the driving task of the vehicle itself. Third, this paper has been enhanced in replay buffer, using more collision and arrival experiences to train the neural network. In this paper, the performance of the original and improved models is evaluated in terms of safety and efficiency. The simulation results show that all three proposed improvement methods can improve performance and achieve better results.