Digital-Twin-Based Deep Reinforcement Learning Approach for Adaptive Traffic Signal Control

Urban vehicle emissions are one of the main contributors to air pollution since most vehicles still rely on fossil fuels, despite the growing popularity of alternative options such as hybrids and electric cars. Recently, Artificial Intelligence (AI) and automation-based controllers have gained attention for their potential use in adaptive traffic signal control. Many studies have been conducted on the application of Deep Reinforcement Learning (DRL) models to reduce travel time in adaptive traffic signal control. However, limited research has been done on adapting traffic signal control to reduce CO2 emissions and fuel consumption in urban vehicles. As such, this work proposes a digital-twin-based adaptive traffic signal control approach that relies on a digital twin of urban traffic network and uses the DRL Multi-Agent Deep Deterministic Policy Gradient (MADDPG) to optimise for reduced fuel consumption and CO2 emission. The system is designed to simulate different traffic scenarios and control strategies, enabling for adaptation in traffic signal adjustments. To assess the effectiveness and applicability of the proposed approach, a quantitative simulation is performed using synthetic and real-world traffic datasets from a multi-intersection network in a neighbourhood in Amman, Jordan, during peak hours. The findings suggest that the DRL approach based on digital twins on synthetic networks can reduce CO2 emissions and fuel consumption even when using a basic reward function based on stopped vehicles.