International Journal of Science, Engineering and TechnologyAuthors: Mr. Sayyed Aamir Hussain
Abstract: This paper presents NudgeLight, a novel traffic control framework that integrates large language model (LLM)-driven psychological behavior modeling with a safety-constrained multi-agent reinforcement learning (MARL) strategy to govern decision-making during the yellow interval at signalized intersections within mixed human–autonomous vehicle (AV) environments. The yellow phase represents one of the most safety-critical and behaviorally sensitive segments of intersection control, where drivers must make rapid and uncertain stop-or-go decisions, frequently resulting in high conflict probabilities. To address this persistent challenge, NudgeLight employs LLM-based cognitive inference to predict heterogeneous human driver intentions under time pressure and dynamically adapts AV and signal policies through safe MARL mechanisms that explicitly enforce collision-avoidance constraints and minimize conflict trajectories among interacting agents. A high-fidelity simulation environment replicating realistic mixed-traffic conditions—including diverse driver archetypes, variable AV penetration rates, and heterogeneous roadway dynamics—was constructed to evaluate the proposed framework. Extensive experimental results demonstrate that NudgeLight substantially reduces surrogate safety conflicts, improves time-to-collision margins, and enhances intersection throughput, while preserving the naturalness and comfort of human driving behaviors. Unlike existing approaches that restrict AV operations to deterministic or conservatively scripted responses, NudgeLight delivers adaptive, cognitively informed, and safety-assured control tailored to real-world behavioral variability. This research provides critical insights for scalable deployment of intelligent, human-centered signal control solutions and contributes to the advancement of safe and harmonious human–AV coexistence in emerging urban mobility systems.
