Practical Learning-Based Approaches for Train Timetable Rescheduling (TTR)
Objective: The project aims to explore practical learning-based approaches for train timetable rescheduling (TTR), and the key focus is its efficiency and practicability (deployable) for TTR under complex environments, for example, unplanned disruptions.
Methodology: The project will propose a knowledge-assisted offline reinforcement learning (KORL) model for online train timetable rescheduling. It combines reinforcement learning with expert knowledge by learning optimal TTR policy from offline data and adapting policies online by interacting with environments. It has three key components including efficient RL algorithms (learn from few samples), adversarial attack training (ensure the robustness of solution models), and online model adoption (the trained model actively adapts to deployment environments).