Integrating a Large Language Model with Reinforcement Learning for Intelligent Traffic Signal Control

Effective traffic light systems are essential to improve traffic flow, fuel efficiency, and public health. Traditional traffic light systems operate on fixed timers that give each direction equal green light time, making them inefficient under varying road environments. This study proposes and evaluates a hybrid adaptive traffic signal control system combining Proximal Policy Optimisation (PPO) reinforcement learning (RL) with a Large Language Model (LLM), providing an agentic extension for emergency scenario handling. A PPO agent is trained in the Simulation of Urban MObility (SUMO) simulator using synthetic, realistic traffic conditions. The standalone PPO agent reduced average halting vehicles per step by 31% and average waiting time per vehicle by 16% relative to the best fixed-timer configuration. The hybrid system achieved comparable regular-traffic performance, with average halting vehicles per step and average waiting time per vehicle both remaining within 3–4% of the standalone PPO agent, while completely eliminating emergency vehicle wait times. These findings demonstrate that a hybrid agent consisting of a PPO reinforcement learning agent and extended with an LLM for emergency response handling can significantly outperform fixed-timer configurations across traffic delay and congestion metrics, while effectively handling critical edge cases.