DeepMind, an artificial intelligence research company founded in 2010 and acquired by Google in 2014, worked with Google Maps to improve estimated arrival times by 20-50% using modern techniques, Supersegments and Graph Neural Networks.
Other than getting more accurate ETAs so you can better plan your day, the research DeepMind did to arrive at such a significant improvement is impressive.
Taking away all of the technical jargon, they figured out how to split up the world's roads into smart roads and have those smart roads use a shared "brain" to figure out where traffic is and where it will be in the future!
You have to admit, that's pretty cool!
- This research's first main idea is the Supersegments (i.e., smart roads), which "consist of multiple adjacent segments of road that share significant traffic volume." Imagine sections of roadway which has its own memory of its past traffic and can predict future traffic.
- The architecture for the Supersegments went through several iterations. Researchers first used existing traffic road segmentation. However, due to each segment being a different size, each Supersegment would need its own uniquely trained neural network (i.e., mucho time and resources). They finally settled on Graph Neural Networks.
- A Graph Neural Network (GNN) is a variation of Reinforced Learning (RL), which is a machine learning root idea like supervised learning (e.x., classification and prediction) and unsupervised learning (e.x., group similar things together).
- With RL, think of a robot, like a Roomba, learning its environment. Other applications include elevator scheduler, drone stabilization, and checkers.
- With the Google Maps problem, the ability for GNNs to generalize the concept of "proximity" allows for predictive power gains, particularly when including adjacent and parallel roads. The model can predict when a traffic jam on one street will spill over into the main street and cause further delays.
There are even more technical specifics, including information on how they made the model production-ready and some discussion about the loss functions.