Reinforcement Learning Driven Policy Optimization for Adaptive Traffic Control Systems

Abstract

Traffic congestion at key intersections in Colombo has been a persistent issue, exacerbated by rising vehicle numbers and the limitations of traditional traffic control methods. Fixed-time signal schedules and manual interventions often fail to adapt to the city’s fluctuating traffic patterns, leading to long queues, delays, and frustration for commuters. This research proposes an adaptive traffic signal control framework for a major intersection in Colombo, addressing these challenges by leveraging real-time traffic data collected through a strategically positioned camera near the junction. The system continuously monitors and assesses traffic flow patterns, using a Long Short-Term Memory (LSTM) network to predict upcoming traffic volumes. A hybrid control algorithm then integrates rule-based congestion management with a Deep Deterministic Policy Gradient (DDPG) agent, optimizing signal timings dynamically to respond to real-time traffic demands. Designed with cost-efficiency, low computational needs, and compatibility with existing infrastructure in mind, this framework aims to improve vehicle throughput, reduce waiting times, and alleviate congestion offering a scalable approach for urban traffic management.