V2I-Enhanced State Representation for Adaptive Traffic Signal Control Using Deep Q-Networks

Abstract

Traffic congestion remains a major challenge for intelligent transportation systems. Vehicle-to-Infrastructure (V2I) communication provides a promising foundation for adaptive traffic signal control by enabling enhanced real-time traffic detection. This paper proposes an adaptive traffic signal control framework that exploits a V2I-enhanced state representation for isolated intersections using a Deep Q-Network (DQN) agent. The agent learns an optimal control policy directly from V2I-derived state information, including queue lengths, traffic signal phase, and green-light duration, to dynamically adjust signal timing and improve traffic efficiency and stability. The proposed framework was evaluated in a custom stochastic traffic simulation environment implemented in Python/PyTorch, trained for 200 episodes, and assessed over 30 independent stochastic runs under low-, medium-, and high-traffic demand scenarios. Experimental results demonstrate that the proposed approach significantly outperforms both heuristic V2I and fixed-time baseline controllers, reducing the mean queue length by 29.8% under medium traffic demand and 58.3% under high traffic demand, while substantially reducing queue variability across stochastic traffic realizations. Furthermore, a comparative evaluation involving DQN, Double DQN, PPO, and A2C revealed no statistically significant performance differences among the reinforcement learning algorithms, indicating that the observed performance gains primarily originate from the proposed V2I-enhanced state representation rather than from a specific reinforcement learning algorithm. These findings demonstrate the potential of V2I-enabled reinforcement learning as a robust framework for adaptive traffic signal control at isolated intersections.