Intelligent Traffic Optimization Using Reinforcement Learning & Agentic AI - Digital Twin Simulation

Abstract

Traffic jam is an unresolved issue in the burgeoning urban centres culminating in the excess of delays, increased fuel usage and high levels of green house gasses. The traditional traffic signal systems are run on a set timing schedule that fails to adapt to the fluctuations that occur in the traffic, thus wasting the road space. In this paper, a Digital Twin-Actuated Intelligent Traffic Optimization System has been suggested, which comprises Proximal Policy Optimization (PPO)-based Reinforcement Learning and Agentic AI supervision module to autonomously operate traffic lights. Microscopic traffic simulation tools are used to create a high-fidelity digital representation of a real intersection of the real world, which can be trained safely and at scale by the learning agent without causing disruption to the real world. The PPO agent receives the real-time traffic condition, queue length, waiting time, traffic density, and signal phase setting and actively decides optimal transition between the phases and green signal length. The development of a novel reward formulation subjects delay reduction, reductions in queues, maximization of throughput, and switching penalties to achievable and stable learning. In order to provide robustness, an Agentic AI supervision layer constantly checks the trends of reward convergence and adjusts hyperparameters, like learning rate, discount factor, entropy coefficient, and batch size. The proposed structure is compared with the conventional Fixed-Time and Max-Pressure control strategies based on the main key performance indicators such as the average waiting, queue length, throughput, and emission estimation. The experimental results indicate that traffic efficiency and environmental sustainability improve significantly, which outlines the scalability and feasibility of digital twin-enabled autonomous traffic optimization systems to be used in smart cities.