So far, machine learning has improved the accuracy of real-time ETAs by up to 50% in cities around the world, the companies said.
It’s hard to remember life before navigation apps and the ability to pop in an address and find out the best route to take, whether traffic is light or heavy, and an estimated travel time and estimated time of arrival (ETA).
While the interfaces are simple to use and often save drivers a lot of time and stress, behind the scenes there is a lot going on to deliver this information quickly.
SEE: TechRepublic Premium editorial calendar: IT policies, checklists, toolkits, and research for download (TechRepublic Premium)
Google Maps helps users navigate over one billion kilometers in more than 200 countries and territories daily by aggregating location data to understand traffic conditions all over the world. This helps determine where traffic is and how that might impact a person’s drive. What it doesn’t do is account for what traffic will look like 10, 20 or even 50 minutes into their journey, explained Johann Lau, product manager at Google Maps, in a blog post.
To predict what traffic will look like in the near future, Google Maps analyzes historical traffic patterns for roads over time. Now, Google Maps has partnered with DeepMind, an Alphabet artificial intelligence (AI) research lab, to improve the accuracy of its traffic prediction capabilities.
“By partnering with DeepMind, we’ve been able to cut the percentage of inaccurate ETAs even further by using a machine learning architecture known as Graph Neural Networks,” Lau said.
So far, this approach has improved the accuracy of real-time ETAs by up to 50% in places like Berlin, Jakarta, São Paulo, Sydney, Tokyo, and Washington, DC, he said. “This technique is what enables Google Maps to better predict whether or not you’ll be affected by a slowdown that may not have even started yet.”
While historical traffic patterns have been reliable indicators for what road conditions could look like, since the start of the COVID-19 pandemic, traffic patterns around the globe have shifted dramatically, Lau said.
“We saw up to a 50% decrease in worldwide traffic when lockdowns started in early 2020,” he said. “Since then, parts of the world have reopened gradually, while others maintain restrictions. To account for this sudden change, we’ve recently updated our models to become more agile—automatically prioritizing historical traffic patterns from the last two to four weeks, and deprioritizing patterns from any time before that.”
SEE: Natural language processing: A cheat sheet (TechRepublic)
Google Maps’ predictive traffic models also help determine driving routes. That’s where alternate route suggestions come in. “If we predict that traffic is likely to become heavy in one direction, we’ll automatically find you a lower-traffic alternative,” Lau said.
To do this, Google Maps looks at a number of other factors, such as road quality: Whether the road paved or unpaved, or covered in gravel, dirt or mud, he said. How direct a road is and whether driving down a highway is often more efficient than taking a smaller road with multiple stops is also taken into consideration, he said.
Two other key sources of information are used in recommending the best route: Authoritative local government data and real-time user feedback, Lau said. Both sources are also used to help understand when road conditions change unexpectedly due to mudslides, snowstorms, or other forces of nature, he added.
The Graph Neural Network model “treats the local road network as a graph, where each route segment corresponds to a node, and edges exist between segments that are consecutive on the same road or connection through an intersection,” according to a DeepMind blog post.
The network models traffic by understanding how cars flow through roads.
“By spanning multiple intersections, the model gains the ability to natively predict delays at turns, delays due to merging, and the overall traversal time in stop-and-go traffic,” DeepMinds said. “This ability of Graph Neural Networks to generalize over combinatorial spaces is what grants our modeling technique its power.”