LLM-Assisted Semantic Stop Embedding for Bus Holding Control

A recent study presents a novel approach utilizing LLM-assisted semantic stop representation aimed at enhancing event-driven bus holding control to reduce bus bunching. This phenomenon negatively impacts service consistency and leads to longer wait times for passengers in high-frequency transit systems. Current reinforcement-learning-based holding controllers depend on immediate operational data or specific stop identifiers, limiting their understanding of the broader operational context and hindering policy application across different routes. The research employs an LLM offline to convert diverse stop data—encompassing physical features, contextual activities, and historical performance—into fixed semantic embeddings. These embeddings are then integrated into a deep Q-learning controller, eliminating the need for real-time LLM processing. Experiments conducted through stochastic simulations, calibrated with data from two bus routes, demonstrate that the proposed method outperforms the best calibrated Daganzo baseline. The paper can be found on arXiv under ID 2608.10207.