RL with Semantic Stop Embedding Mitigates Bus Bunching.

Bus bunching is a common problem in high-frequency transit systems, leading to irregular service and increased passenger waiting times. Existing reinforcement learning (RL) controllers for bus holding often rely on limited operational data or route-specific identifiers, which restrict their ability to understand the full context of stops and hinder policy reuse across different routes. This research proposes a novel approach that enhances RL-based bus holding control with an LLM-assisted semantic stop representation. An LLM is used offline to convert diverse stop information—including physical attributes, surrounding activities, and historical operational data—into fixed semantic embeddings. These embeddings are then integrated into a deep Q-learning controller without requiring real-time LLM inference. Simulations calibrated with real-world data demonstrated significant improvements: the semantic controller reduced headway variability by 32%, bunching events by 69.2%, and passenger waiting time by 24% compared to the best baseline. The semantic information proved superior to simple route-specific identifiers, offering a better trade-off across control objectives. Furthermore, while zero-shot transfer showed limited immediate generalization, warm-start fine-tuning accelerated learning for new routes, suggesting potential for adaptation-based policy reuse.