A Deep Learning, Model-Predictive Approach To Neighborhood Congestion Prediction And Control
Lecture | August 29 | 11 a.m.-12 p.m. | 212 O'Brien Hall
Sudatta Mohanny, UC Berkeley
The talk explores a technique for effectively representing the congestion state in a neighborhood for the purpose of short-term predictions through a simple scoring mechanism based on the Macroscopic Fundamental Diagram (MFD). Network state signals emanating from a larger region-wide network are then utilized to predict this score using a deep learning (LSTM) framework, which is further enhanced by Spectral Graph Theory based feature learning (Graph-CNN). The model accuracy is found to significantly outperform several baselines (Nearest Neighbors, Holt-Winters and LSTM models) for both a simple toy network and a larger network representing the San Francisco Bay Area, while proving to be robust against noisy or missing data sources and also providing information about the causes of congestion through a Neural Attention based framework. Finally, the model predictions and attentions are shown to be useful in designing a novel app-based congestion pricing strategy which not only leads to significant reduction in delays, but also allows for much lower toll rates than other naive strategies, thus implying much lower discomfort to travelers.
Sudatta Mohanty is a PhD candidate in Transportation Engineering, jointly advised by Prof. Michael Cassidy and Prof. Alexey Pozdnukhov. His research interests are in Applied Machine Learning, Traffic Flow Theory, Graph Theory and Optimization. He has had significant industry experience through internships at Apple, Microsoft Research and Sidewalk Labs.