Waymo uses evolutionary competition to improve its self-driving cars

The procedure of training self-driving car AI is rarely capable when you require either using a massive amount of computing power for training systems in parallel or else having researchers who spend years manually weeding out bad systems.

But Waymo may have a smarter approach and using the similar principles that guide evolution. Waymo has partnered with DeepMind on a ‘Population Based Training’ technique for pedestrian detection that has the most excellent neural networks advance much similar to lifeforms do in natural selection, saving time and effort as well.

The approach commonly has the networks competing against each other, with weaker examples being substituted by stronger ‘progeny’ that are copies of the good performing networks with a little tweaked parameters. This automatically gets free of the poorer performing networks even as saving Waymo from having to retrain networks from scratch they have previously inherited know-how from their parents.

There is a risk that the technique is focused very much on improvement which are on short term basis. For fighting this Waymo formed several ‘niches’ where neural networks challenged to each other in sub groups for getting good results although preserving assortment that can be better-suited for driving conditions of real world.

Moreover, when applied to pedestrian detection the consequences were promising. The PBT approach dropped fake positives by 24%, even though it took partially as much time. The experiment went so good that Waymo has even been applying PBT across other models. That, in turn, promises self-driving cars that can improved cope with the difficulties of driving and hence avoid collisions.