Scaling Real Time Traffic
- Source
- uber.com
- Author
- Uber Engineering 2
- Date
Context
Uber's DeepETT traffic forecaster, described by Uber Engineering, initially exposed a problem: when Uber retrained its downstream arrival-time models on the new forecasts, final arrival-time accuracy decreased even though road-segment forecasts had improved. Small systematic errors changed over time and accumulated when segment forecasts were combined into a trip.
The team added continuous , comparing forecasts with recent observed travel times and applying corrections. This kept the forecasts consumed by downstream arrival-time models stable while retaining the new model's useful traffic information. Uber reports that DeepETT as a whole improved long-trip arrival-time accuracy by 6% in a global A/B test. In Uber's account, the gains in forecast quality could reach riders only if calibration was maintained after training, not just at training time.
- calibration — How well a model's confidence matches reality — a calibrated model saying "90% sure" is right about 90% of the time.
- transformer — The neural network architecture behind modern AI models, built on attention — letting every word directly consider every other word in parallel.
Checking sign-in…
Loading comments…