LLM-Enhanced AI Improves Adaptive Traffic Control

Managing urban traffic signals is a complex challenge, requiring real-time adaptation to balance multiple objectives like throughput, delay fairness, and emergency vehicle priority. Existing reinforcement learning (RL) methods often struggle with fixed objectives, dynamic priority changes, and generalizing across geometrically similar intersections. This paper introduces SIGMA, an innovative RL framework designed to address these limitations. SIGMA (Symmetry-aware, Intelligent, Geometric, Multi-objective Adaptive traffic control) integrates a large language model (LLM) to enable adaptive objective tuning. This allows the system to convert natural-language emergency commands into priority vectors for a multi-objective actor-critic controller, eliminating the need for manual reward engineering. Furthermore, rotational augmentation enhances the system's transferability across different four-way intersections, while an offline-to-online learning approach ensures stable initialization and gradual adaptation to changing traffic conditions. The framework defines and validates reliability properties, including emergency service levels and graceful degradation under LLM failures. Evaluated in SUMO on real-world Kolkata intersections, SIGMA significantly outperforms traditional controllers (fixed-time, actuated, DQN) by reducing average and emergency waiting times, decreasing queue lengths, and increasing throughput. Ablation studies confirm its robustness to component failures and geometric rotations, positioning SIGMA as a reliable, language-guided, and multi-objective solution for advanced traffic management.