SmartSwitchLight: Interpretable multi-agent reinforcement learning for adaptive traffic signal control via generative artificial intelligence based supervisory decision making

Urban traffic signal control must respond to rapidly changing demand caused by commuting patterns, special events, and non-recurrent congestion, yet conventional fixed-time, actuated, and single-agent reinforcement learning (RL) controllers often struggle to adapt across diverse operating regimes. This paper presents SmartSwitchLight, a generative artificial intelligence (GenAI) supervised multi-agent traffic signal control framework for adaptive urban traffic signal control. The implemented GenAI combines multiple specialized deep reinforcement learning (DRL) controllers, each trained for a distinct signal phase configuration, with a large language model (LLM) based meta-controller that selects the most suitable controller according to real-time traffic conditions, profile-based forecasts, historical performance, and contextual information. Rather than directly issuing low-level signal timing actions, the meta-controller performs high-level policy selection and provides interpretable supervisory reasoning for each decision. The proposed system is evaluated in a high-fidelity microscopic simulation environment under a realistic full-day demand profile with strong directional peaks and event-driven fluctuations. Results show that SmartSwitchLight substantially outperforms fixed-time, actuated, and standalone RL controllers. The quarter-hourly switching configuration reduces mean total intersection delay by 94.7% relative to the best fixed-time controller, 87.9% relative to the best actuated controller, and 83.2% relative to the standalone RL controller. The hourly switching configuration achieves corresponding reductions of 92.4%, 82.7%, and 76.0%, respectively. These findings demonstrate that combining specialized RL policies with GenAI-based supervisory decision making can significantly enhance adaptability, robustness, and interpretability in adaptive urban traffic signal control.