Google’s self-driving car team just released its January report, which highlights the role played by its simulator in improving its driving algorithms.
With our simulator, we’re able to call upon the millions of miles we’ve already driven and drive those miles again with the updated software. For example, to make left turns at an intersection more comfortable for our passengers, we modified our software to adjust the angle at which our cars would travel. To test this change, we then rerun our entire driving history of 2+ million miles with the new turning pattern to ensure that it doesn’t just make our car better at left turns, but that the change creates a better driving experience overall.
And the simulator isn’t not limited to what the car has already seen:
We can also create entirely new scenarios in our simulator, allowing us to concentrate on perfecting a particular skill. For example, to test our car’s performance in a three car merge, we will create thousands of variations of this situation (each car travelling at different speeds, and nudging to merge at different times) and then test that our car drives as intended each time.
To me, this is one of the coolest parts of machine learning. Without actually going out and getting new data, which can be expensive and slow, we can use data that we already have, and warp it to create lots of new data, which rapidly improves the learning rate of machines.
Originally published at www.davidincalifornia.com on February 3, 2016.
David Silver 