Enhancing Traffic Signal Management through Hybrid Models
Author : Dr. T Suvarna Kumari, A Mayaank and Nama Sricharan
Abstract :
Modern-day urban traffic management is turning complex with increasing traffic volumes and intricate intersection dynamics, demanding advanced control methodologies. Traditional solutions in Traffic Signal Control (TSC), such as the Webster technique and Self-Organizing Traffic Light Control (SOTL), are severely limited since they rely on predesigned rules and assumptions, thereby showing comparative ineffectiveness in dynamic environments. While existing Reinforcement Learning-based TSC systems are more flexible, they still have significant challenges to overcome in imperfect observation handling-like degraded communication- and rare events not covered by the designed reward functions. In this respect, this project proposes a new framework that integrates RL with LLMs to improve the strength of the process of traffic signal control. It operates in two major steps: the first one involves an RL agent optimizing traffic light timing based on the recorded traffic flow data; the second invokes an LLM evaluating and further refining the decisions of the RL agent by the prompt engineering to obtain the optimal outputs. This integrated approach offers a much more adaptive and robust solution to TSC, with a view toward tackling the complexities of modern urban traffic. While this work presents a pointing concept of combining LLMs with RL. for TSC, this is just the beginning in the exploration of its applications. Further work may include increasing the capacity of the framework for handling more challenging intersections and other situations.
Keywords :
Traffic management, TSC, SOTL, vehicle detection system etc.
