Revisiting Traffic Forecasting from a PINN Perspective

Arian Prabowo, Flora D. Salim

Primary Area: applications to robotics, autonomy, planning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: traffic forecasting, pinn

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: We use the lens of symmetry to compare data-driven (DL) and model-driven (PDE/CFD) approach to traffic.

Abstract:

This paper revisits various traffic forecasting models and provides a fresh perspective by examining them through the lens of Physics-Informed Neural Networks (PINNs). Instead of proposing new models, our approach focuses on identifying common methods within existing models and elucidating why these methods are effective from a PINN perspective.

We explore the concept of symmetry in traffic forecasting models, particularly in deep learning models and classical model-driven approaches that simulate traffic as fluids using partial differential equations (PDEs). We observe that deep learning models often exhibit spatial symmetry in their main backbone, with the exception of node embeddings. In contrast, classical PDE-based traffic models feature spatial symmetry in the PDE but not in the boundary conditions. This insight suggests that node embeddings encode critical boundary conditions in these models.

Additionally, we investigate the relationship between adaptive adjacency and graph re-wiring, shedding light on the dynamic nature of traffic network representations. By drawing parallels between these concepts, we offer valuable insights into how traffic systems adapt to changing conditions.

In conclusion, this paper presents a unique perspective on traffic forecasting models, emphasizing the role of symmetries and boundary conditions. It envisions a future where deep learning models harness the full capabilities of classical simulations, enabling the exploration of counter-factual questions. Such advancements hold great promise for traffic planners, managers, and engineers seeking more comprehensive and effective solutions for urban transportation management.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 5110