The world is being quietly rearranged by people who write very long documents.


The title they went with Interpretable long-term traffic modelling on national road networks using theory-informed deep learning Noisy translates that to

New method predicts highway traffic with theory and AI combined


Researchers created a machine learning system that predicts how much traffic will use specific highway segments years into the future, by combining economic theory about how people choose routes with neural networks trained on eight years of UK road data. This matters because transport planners need accurate long-term forecasts to decide where to build roads or improve capacity, but existing methods either work well on paper but require endless manual tuning, or learn patterns from data but can't explain why or work in new regions.
For decades, traffic forecasting has used economic models that make strong assumptions about human behavior but struggle with real-world complexity, or black-box machine learning that predicts accurately but can't tell you why or transfer to new roads. This paper shows those don't have to be separate — you can embed transport economics directly into a learning system, making it both accurate and intelligible. That matters because transport infrastructure is expensive and long-lived; planners need to trust the forecast enough to bet decades of funding on it, and currently they have to choose between understanding the model or trusting its accuracy.

If you insist
Read the original →