Research on Lane Changing Decision of Intelligent Driving Vehicles Based on Improved Deep Reinforcement Learning

With the rapid development of autonomous driving technology, decision-making capability in complex road environments has become a crucial research focus. In highway scenarios, lane-changing is a key behavior that directly impacts driving safety and efficiency. This paper proposes an improved lane-changing decision-making method based on the Deep QNetwork (DQN) reinforcement learning algorithm. To address the problem of Q-value overestimation in traditional DQN, we adopt the Double DQN (DDQN) framework, which separates the selection and evaluation of actions. Experiments are conducted using the real-world high D highway trajectory dataset, where the state space is composed of 13 features including vehicle velocity, acceleration, distance, and relative speed to surrounding vehicles. The action space consists of three lane-change decisions: left, stay, and right. Experimental results show that the improved DDQN model significantly outperforms the original DQN in terms of accuracy, F1-score, collision rate, and success rate, thereby enhancing the safety and reliability of autonomous lane-changing decisions. The proposed method offers an effective reinforcement learning-based solution for intelligent driving decision-making systems.