Google Maps is one of the most widely used products in the company, and the ability to predict upcoming traffic jams is essential for many drivers. Every day, Google says, more than 1 billion kilometers of road are run with the help of the app. But, as described in the search giant a Blog post Today, more precise thanks to its features Machine learning tools from Deepmind, The London-based AI Lab, owned by Alphabet, Google’s parent company.
In the blog post, Google and Deepmind researchers explain how to take data from different sources to estimate traffic flows and feed it to machine learning models. This data includes live traffic information collected anonymously from Android devices, historical traffic data, information such as speed limits and construction sites from local governments, and the quality, size and direction of any road. So, in Google’s estimates, paved roads hit the unpaved ones, but the algorithm determines that it takes longer to take the motorway than navigating multiple closed streets.
All of this information is fed into neural networks created by Deepmind, which select patterns in the data and use them to predict future traffic. Google says its new models have improved the accuracy of Google Maps real-time ETAs by up to 50 percent in some cities. It also states that the data it uses will need to be changed to make these estimates following subsequent changes in COVID-19 prevalence and road use.
“When lockdowns began in early 2020, there was a 50 percent drop in global traffic,” wrote Johann Law, Google Maps product manager. “For this abrupt change, we have recently updated our models to be more active – automatically prioritizing historic traffic policies over the past two to four weeks and draining samples anytime before.”
Adjacent street groups that share traffic volume – Google works by dividing maps into so-called “supersegments”. Each of these is paired with a personal neural network that makes traffic predictions for that sector. It is not clear how big these supersegments are, but Google suggests that they have “dynamic sizes”, that they change like traffic, and that each attracts “terabytes” of data. The key to this process is the use of a special type of neural network called a graph neural network, which Google says is uniquely suited for processing this type of mapping data.