Aug 22, 2017 · 1 min read
Thanks! You know, I did consider adding a “supervised” element to the Autopilot’s reinforcement learning algorithm. Passengers could manually report when “bad things” happened, like if the Autopilot ran over a piece of shredded tire. Presumably, if enough people participated, the fleet would learn how to avoid hitting shredded tires (at least most of the time).
This type of passenger-directed learning would be crowdsourced, kind of like how Waze depends on people manually reporting the current conditions of the road. I do think it’s a really cool idea, but I think setting up an entirely unsupervised system would create more good data to work from, and would result in fewer malicious/false reports.