Paper ID: 2408.15751

Adaptive Traffic Signal Control Using Reinforcement Learning

Muhammad Tahir Rafique, Ahmed Mustafa, Hasan Sajid

Traffic demand is continuously increasing, leading to significant congestion issues in major urban areas. Constructing new infrastructure is a potential solution but presents a substantial financial burden on national economies. An alternative approach involves optimizing existing traffic networks through the dynamic control of traffic signals at intersections. Recent advancements in Reinforcement Learning (RL) techniques have demonstrated their capability to address the complexities associated with traffic congestion. In this paper, we propose a solution to traffic congestion using reinforcement learning. We define the state as a scalar representing the queue length, demonstrating that the algorithm can effectively learn from this simplified state representation. This approach can potentially reduce deployment costs by minimizing the number of sensors required at intersections. We have developed two RL algorithms: a turn-based agent, which prioritizes traffic signals for the intersection side with higher traffic, and a time-based agent, which adheres to a fixed phase cycle, adjusting the phase duration based on traffic conditions. To assess the performance of these algorithms, we designed four distinct traffic scenarios and computed seven evaluation metrics for each. Simulation results indicate that both algorithms outperform conventional traffic signal control systems.

Submitted: Aug 28, 2024