I find it intriguing that DeepMind’s AlphaStar used supervised imitation learning on a dataset of 500,000+ human-played games. With supervised imitation learning alone, it was able to beat the StarCraft Elite AI bot 95% of the time, and get an estimated MMR that would put it in the Gold league — where the middle third of ranked human players are. So just with imitation learning, AlphaStar was able to get to roughly median human performance (or at least we think so; I don’t know if it beat any Gold-ranked human players).
The Information reported (https://www.theinformation.com/articles/what-makes-teslas-autopilot-different) that Tesla is using supervised imitation learning to “directly predict the correct steering, braking and acceleration” based on the state/action pairs observed in human driving. Should we feel optimistic that what (apparently) worked for StarCraft will work for autonomous cars?