Nearly 13 years ago, Google Maps started providing navigation data for drivers, along with estimated travel times and traffic density estimates.
To bring the technology to the next level, researchers from Google and Alphabet’s AI research lab DeepMind managed to improve the estimated time of arrival (ETA) accuracy by up to 50% in locations such as:
- São Paulo
- Washington DC
The improvement in accuracy was made possible thanks to graph neural network, a machine learning technique.
Google Maps product manager Johann Lau explained that historical traffic patterns and aggregate location data are utilised in the new way that traffic estimates are calculated.
Their current ETA predictions have a 97% accuracy.
Using graph neural networks allows Google Maps to incorporate relational learning biases.
Based on this, modelling the connectivity structure of real-world road networks is made possible.
DeepMind researchers also claim that they experimented with adjacent roads that are not part of the main road.
The new model has the capacity to predict delays at turns, delays that occur due to merging, and stop-and-go traffic.
Graph neural processing takes all of this into account and provides accurate estimates.
The machine learning model of Google Maps has also been adjusted to take into account the global traffic pattern changes that have emerged since the beginning of the COVID-19 pandemic.
Lau reports that the model assigns a higher priority to traffic patterns emerging from the last two to four weeks compared to the ones that surfaced before that time frame.
Google Maps also determined driving routes based on:
- Predictive driving models
- Real-time feedback from users
- Authoritative data from local governments
If Google Maps determines that one route is traffic-heavy, it will suggest a new, lower-traffic alternative.
During calculations, road quality is also taken into account.