Google Maps uses DeepMind tools to predict arrival times
Google said that Google Maps has been enhanced with machine learning tools from DeepMind, a London-based artificial intelligence lab owned by Alphabet.
Researchers at Google and DeepMind explained how they take data from various sources and inject it into machine learning models to predict traffic flows. This data includes live traffic information collected anonymously from Android devices, history of traffic data, information such as speed limits and road works from local authorities, and the quality, size, and direction of the road chosen.
According to Google estimates, paved roads outperform unpaved roads in speed, and the algorithm decides that sometimes it is better to drive longer on a motorway than to navigate several winding streets. All this information feeds into neural networks developed by DeepMind, which predict future traffic from data patterns.
Google noted that new training models have increased the accuracy of Google Maps in real-time by up to 50% in some cities. The data itself for the forecasts had to be updated after the outbreak of COVID-19. The company explained that it saw a drop in global traffic of up to 50%.
Models work by dividing maps into super-segments, which are the joins of adjacent streets. Each of them is connected to an individual neural network that makes traffic predictions for that sector. It is unclear how large these super-segments are, but Google notes that they have "dynamic parameters," meaning they change with traffic, and each uses "terabytes" of data.
The forecasting process uses a graph neural network, which Google claims is particularly well suited for processing this kind of map data.
Recently, Google added statistics on the incidence of COVID-19 in the locality of interest to the map service. In addition to information about hotels or attractions, the user will see data regarding restrictions related to COVID-19, as well as the resumption of flights to a specific destination.