A Deep Q-Learning Approach to Improve the Effects of Novice Drivers’ Driving Behavior at a Signalized Intersection

This study presents a deep Q-learning optimization framework to improve the effects of novice drivers’ behavior on signalized intersection performance. The driving behavior of 70 novice drivers was collected at a signalized intersection using drone surveillance and camera systems. Key driving parameters, including standstill distance (CC0), perception-reaction times, acceleration-deceleration characteristics (CC8), and speed profiles, were analyzed via YOLOv8-based image processing. A deep Q-learning algorithm was developed with a state space containing vehicle inputs and signal timing states, where the reward function minimizes vehicle delays while maximizing travel speeds. The agent learns optimal signal timing policies through continuous interaction with a calibrated simulation environment of novice driver behaviors, and the framework was tested at various novice driver rates (0%–50%). The optimization achieves 5% reduction in vehicle stops, 8% reduction in delays, and 10% improvement in queuing delays. This study contributes to the literature by combining real-world novice driver data with deep Q-learning optimization for signal control and demonstrates that reinforcement learning can effectively reduce the impact of novice drivers on intersection performance.