Deep Reinforcement Learning based Joint Optimization of the Traffic Signal and Autonomous Vehicles Considering Dynamic Uncertainties

Abstract

The increasing availability of sensor data and advancements in deep reinforcement learning (DRL) offer promising opportunities for enhancing traffic management efficiency. However, most existing DRL methods for traffic management concentrate solely on enhancing signal controls. To address this limitation, we propose a novel DRL-based joint optimization method that integrates traffic signal control with the longitudinal control of connoted autonomous vehicles (CAV ). This method enhances overall traffic efficiency by synchronizing signal control with CAV behaviour. To alleviate dynamic uncertainty problems in joint optimization, this method leverages temporal prediction networks, ideal for anticipating future traffic states, and fuzzy neural networks, adept at handling information with uncertainties and errors, to extract useful and accurate traffic representations in mixed-traffic environments. A case study across various isolated intersections is conducted to validate the effectiveness of the proposed method. The results demonstrate that our method outperforms state-of-the-art methods on both synthetic and real-world datasets.