You can see how radical and worldwide AI implementations really can be, if you look at the collaboration between Alphabet subsidiary and AI research company, Deepmind and Google Maps. For the first time in thirteen years, historical traffic data of Google Maps was no longer a perfect comparison framework due to Covid-19 which has resulted in strongly deviating traffic patterns. By using Graph Neural Networks, Google can now determine your travel time even better than before!
Traffic depends on many parameters
Every Maps user will use the ETA (the expected time of arrival) to record his departure time for a trip. Of course, this ETA is not stable. First and foremost, several roads lead to Rome. And the shortest route is not always the fastest. And purely on the basis of maximum permitted speeds and delays due to traffic lights, you can perfectly determine a theoretical driving time, but unfortunately you are not alone on the road and your travel time in heavy rain showers will not be comparable to those in sunny weather. And of course you also have road works and events that can influence your theoretical travel time. So many parameters to take into account.
One way to calculate travel time is to use historical data. For example Mondays versus Sundays. At eight o’clock in rush hour versus ten o’clock. In January versus August. But in any case, the past only gives a prediction and not necessarily the correct one. Accidents or Corona. They all have their impact. So Google is increasingly relying on real-time data. Lots of people walk and drive around with cell phones, providing a lot of useful information. How many people are driving in the same direction on the same track. What speed are they currently driving at. Because, based on the capacity of that track, you can predict when the traffic will have an impact on the average speed, when the traffic jam will occur and what impact this will ultimately have on your travel time. Even if you pass that point only half an hour later.
And, of course, there are a range of other sources of real-time information. Weather reports, traffic information, government alerts.
Google has been doing all these things for a while, but now they have tinkered with the calculation models with the help of Deepmind. And with impressive results as you can see in the following image:
Bron: Deepmind.com ©
Previously, the predictions in terms of travel time were more than OK (on average more than 97%), but especially in busy locations such as world cities, those figures were a lot lower. But by using Graph Neural Networks, Google can now predict your travel time with greater certainty in busy locations.
But that’s not the only thing that has changed about the model. Covid-19 has resulted in strongly deviating traffic patterns so that the past no longer offered a perfect comparison framework for the first time in thirteen years.
In the new environment, Google has therefore made the models more flexible so that they can respond better to changed circumstances. The new models work by splitting maps into what Google calls “super segments”: clusters of adjacent streets that share traffic volume. Each of these are linked to an individual neural network that makes traffic predictions for that sector. Those super segments have Google’s “dynamic sizes,” which suggests that they change in content as traffic changes. And that ultimately led Google to the following approach for calculating your ETA.
Bron: Deepmind.com ©