Alphabet owns both DeepMind and Waymo, and the two subsidiaries are working together to train AIs that operate Waymo self-driving cars. The way the duo are training the AI is using the technique that DeepMind calls population-based training (PBT) that was developed and used previously by DeepMind for training the video game algorithms used to defeat humans in StarCraft II. The same technique was also used to train AIs to play Quake III Arena.
PBT takes inspiration from biological evolution and can speed up the selection of machine-learning algorithms and parameters for a particular task. PBT does this by having the AI choose the machine-learning algorithms and parameters available for a specific task from the ones that perform the task the most efficiently in an algorithmic population.
The team believes that if PBT can train AI agents to play StarCraft II, PBT will also be able to train neural networks to handle the different decisions required to maintain safety when driving. This sort of training for self-driving vehicle AIs is expected to work very well. The team says self-driving cars need to be retrained and recalibrated as the vehicle collects new data and as they are deployed in new environments. The more flexible an AI is, the more useful it will be in different contexts.
Teams at Waymo are using PBT to improve the development of deep-learning code that is used to detect lane markings, vehicles, pedestrians, and to verify the accuracy of labeled data that is fed to other machine-learning algorithms. According to researcher Yu-hsin Chen, using PBT has reduced the computational power needed to retain a neural net by about half and has doubled or tripled the speed of development. Waymo's team also says that the process involved in using PBT for training an AI makes it easier to understand how a deep-learning algorithm has been optimized and changed over time as it produces something similar to a genealogical tree.