Google Maps uses DeepMind’s AI tools to predict the arrival time

Google Maps is one of the organization’s most generally-used products, and its capacity to predict future traffic jams performs it necessary for many drivers.

Each day states Google, more than 1 billion kilometers of road is driven with the app’s help. But, as the search titan reveals in a blog post today, its features have got more reliable thanks to machine learning tools from DeepMind, the London-based AI lab owned by Google’s parent company Alphabet.

In the blog post, Google and DeepMind researchers describe how they take data from multiple sources and feed it into machine learning models to predict traffic flows.

This data adds live traffic information gathered anonymously from Android devices, historical traffic data, information like speed limits and construction sites from local governments, and also parts like the quality, size, and direction of any given road.

So, in Google’s views, paved roads beat unpaved ones, while the algorithm will determine it’s sometimes faster to take a long stretch of a motorway than navigate various winding streets.

All this data is loaded into neural networks designed by DeepMind that choose out patterns in the data and apply them to predict future traffic.

Google tells its new models have increased the accuracy of Google Maps’ real-time ETAs by up to 50 percent in some cities. It also records that it’s had to improve the data it manages to make these predictions following the outbreak of COVID-19 and the following change in road usage.

“We observed up to a 50 percent reduction in worldwide traffic when lockdowns started in early 2020,” writes Google Maps product manager Johann Lau. “To consider for this sudden turn, we’ve lately updated our models to become more agile automatically prioritizing historical traffic models from the last two to four weeks, and deprioritizing patterns from any time before that.”

The models run by distributing maps into what Google calls “super segment” clusters of nearby streets that partake traffic volume. Each of these is joined with a single neural network that makes traffic predictions for that sector.

It isn’t apparent how large these super segments are, but Googles notes they have “dynamic sizes,” implying they change as the traffic does, and that each one draws on “terabytes” of data.

The key to this method is the use of a particular type of neural network known as Graph Neural Network, which Google states are especially well-suited to treating this sort of mapping data.