Ongoing Project:

Objective: generating origin-destination (OD) matrix under given conditions based on history OD data

Method: Coarse-to-fine generation via HRL + energy scores as similarity metric

• Developed a hierarchical reinforcement learning framework that can automatically decompose spatio-temporal generation tasks into long- and short-term goals;
• Implemented a goal-directed hierarchical Decision Transformer with 2D coordinate awareness to explore mixed continuous/discrete action space.
• Deployed in various application scenarios, such as epidemic forecasting and urban planning.

Project

DOI