Waymo uses evolutionary competition in order to improve the AI system implemented in its autonomous vehicles. In order to implement evolutionary competition, Waymo partnered with Deep Mind to create "Population-Based Training," which improves pedestrian alerting by using neural networks. Evolutionary competition works by having neural networks compete in order to find out which one is better. The network that does not perform as a good then gets replaced by the one that performed better.
The initial results of the program indicated success. Pedestrian detection rates performed better after the evolutionary competition was implemented, instances of false positives decreased by 24 percent. This drop occurred in approximately half the time it normally would using previous methods.
Concerns about diversity in neural networks were addressed when Waymo created "niches," which are subgroups.
Image Credit: Waymo