Optimizing Traffic Flow in Sri Lanka Using Reinforcement Learning-Based Traffic Light Control Approach

Abstract

Traffic congestion in urban areas has become a significant challenge, particularly in developing countries such as Sri Lanka, where conventional fixed-time traffic signal systems are unable to adapt to dynamic traffic conditions. This research aimed to develop an adaptive traffic signal control system using reinforcement learning to improve traffic flow efficiency.The study primarily focuses on an Advantage Actor-Critic (A2C) reinforcement learning agent for adaptive traffic signal control. The A2C model was trained to optimize traffic flow by utilizing key state inputs such as vehicle waiting time, queue length, and rainfall conditions. Real-time traffic data required for these inputs were obtained using a YOLO-based vehicle detection model, which provided vehicle counts and queue length information. The system was trained and evaluated using a combination of real-world traffic data and simulated environments in the SUMO platform.The results showed that the proposed approach reduced vehicle waiting time and queue length compared to traditional fixed-time control methods. The reinforcement learning agent demonstrated the ability to learn improved signal timing strategies through interaction with the environment.Overall, the findings indicate that reinforcement learning-based adaptive traffic control can enhance traffic efficiency and provide a promising solution for managing congestion in complex urban environments.