Event-Triggered Preference Adaptation with Large Language Model-Guided Conditional Reinforcement Learning for Pedestrian-Aware Traffic Signal Control

Abstract

Traffic signal control at mixed vehicle-pedestrian intersections requires balancing vehicle efficiency, pedestrian service, and operational stability. Existing reinforcement learning methods often rely on fixed reward preferences or unconstrained signal actions, which may limit their adaptability and interpretability under dynamic conflict conditions. To address this issue, the paper proposes an event-triggered large language model-guided conditional reinforcement learning framework. In this framework, the reinforcement learning agent optimizes green duration under a fixed cyclic phase sequence, while the large language model acts as a high-level preference adapter rather than a direct signal controller. A conflict-aware event trigger is designed using pedestrian queue, pedestrian waiting time, and vehicle queue to activate the large language model only under critical states. The generated preference vector guides a preference-conditioned dueling double deep Q-network to balance vehicle efficiency, pedestrian-oriented fairness, and stop-and-go stability. Microscopic simulation experiments in Vissim show that the proposed framework achieves more balanced performance across reward, average delay, stop frequency, pedestrian waiting, and preference-adaptation behavior. Compared with fixed-time control, standard reinforcement learning, and pressure-based control, the proposed method provides stronger adaptability and more interpretable multi-objective decision-making. The results indicate that event-triggered LLM-guided preference adaptation is an effective strategy for pedestrian-aware traffic signal control.