Optimizing Traffic Signal Control in Mixed Traffic Scenarios: A Predictive Traffic Information-Based Deep Reinforcement Learning Approach

The rapid advancement of Connected Autonomous Vehicles (CAVs) is a driving force in the evolution of smart cities and Intelligent Transportation Systems (ITS). This has spurred extensive research in both fields, with a significant focus on vehicle-to-infrastructure (V2I) communication. Deep reinforcement learning is emerging as a popular method in this realm. However, current literature shows a significant gap in exploring the dynamics of traffic flow information for traffic signal control in a mixed traffic environment. Our research addresses this by introducing a predictive traffic information module. This module leverages historical traffic flow data to discern patterns at intersections, enabling proactive traffic signal control by anticipating future traffic states. Alongside this, we developed a reward function where agents, consisting of both traffic signals and CAVs, collaborate towards collective rewards. This strategy not only optimizes traffic signal control but also yields greater environmental benefits. Our experiments indicate that our method outperforms standard benchmarks at an isolated intersection, improving traffic efficiency and reducing environmental impacts by over 20% and 18%, respectively.