The rapid increase in urban vehicle numbers has intensified traffic congestion and safety challenges. Unsignalised intersections pose significant difficulties for autonomous vehicle decision-making. To enhance decision efficiency and safety in such scenarios, this study proposes a decision optimisation method for autonomous driving at unsignalized intersections. The approach first employs a generative pre-trained transformer (GPT) to learn complex interactive behaviour patterns from driving data and acquire prior knowledge. This prior knowledge is then used to initialise the policy network of a deep reinforcement learning (DRL) agent, specifically deep q-network (DQN), which is further optimised through interaction within a simulated environment. This framework aims to combine the powerful sequence modelling capability of GPT with the goal-directed optimisation strength of DRL. Experimental results demonstrate that the proposed method achieves superior performance: The median safe distance reaches 19.58 m (maximum 32.50 m, minimum 8.46 m), the collision rate is as low as 1.07%, and the success rate exceeds 98%. Compared to baseline methods, the proposed approach significantly improves decision-making efficiency and safety for autonomous vehicles at unsignalised intersections, validating its effectiveness.
Bojun Liu (Tue,) studied this question.