Episodes-Based Traffic Signal Control: A Deep Reinforcement Learning Approach With Fluid-Dynamic Simulation

Traffic congestion remains a critical issue in urban traffic networks, leading to increased fuel usage, emissions, and frustration among commuters. This project suggests an AI- driven traffic optimization using the integration of Reinforcement Learning (RL) and fluid-based traffic flow simulation. Taking the Deep Q-Network (DQN) algorithm, the system trains an intelligent agent to learn dynamically adapting traffic light status towards minimizing road congestion. The road is modeled as a grid one-dimensional space in which the car flow is simulated by Taichi, an efficient computer library that simulates cars as a fluid to compute densities accurately. Initial traffic is imported from a density database, which represents car distribution on segments. For every time step, the agent observes significant features such as average, max, and min traffic densities, selects traffic light actions, and receives feedback according to the congestion status of the system. The simulation displays the instantaneous impact of changing lights, and the agent learns to avoid traffic congestion through continuous interaction. The final outcomes are visualized by heatmaps and animations to depict flow pattern improvement. This project describes how AI may control traffic systems autonomously and present a workable, scaleable solution for real-world city mobility problems.