Optimizing Traffic Signal Phases in Mixed Traffic Flow: A Deep Q-Learning Approach

This work explores the optimization of traffic signal phases at a single intersection under mixed traffic conditions, focusing on varying market penetration rates of Connected and Automated Vehicles (CAVs). Utilizing a Deep Q-Network (DQN) approach, the research dynamically adjusts traffic signal phases in response to changes in CAV penetration, aiming to improve traffic flow and alleviate congestion. The results demonstrate that higher CAV penetration rates in mixed traffic scenarios lead to enhanced rewards, reduced delays, fewer emergency braking incidents, and shorter queue lengths, as evidenced by systematic performance improvements across episodes. Furthermore, the analysis of action selection and Q-value differences reveals the model’s preference for specific actions and varying levels of decision-making certainty. The distributions of Q-values underscore the influence of CAV penetration on the stability and predictability of the traffic management system. These results highlight the potential of DQN-based traffic signal control to significantly enhance intersection efficiency in environments with diverse levels of CAV adoption.