Hacker News Reader

Great observations and references @jpfr.

Learning for the AI does not have to be from real world experiences only. Simulated/controlled emergency situations would help as well! Further even if the 2K lines of code stretches a bit more to deal with unknown situations that isn't so bad either

View on HN

But this is a fundamental problem: The learning approach might need 100s of examples of drivers reacting to a bicycle on a sidewalk while turning right into a parking lot to get the right training input. Or perhaps it can learn from examples of bicycles and sidewalks and driveways to do the right thing. The point is, there are millions of edge cases, so getting examples of them all for training or verification is a very large task. The alternative is to build a more general world model where it's possible to work from the other direction and gain confidence that yes, the car senses all other obstacles correctly, and yes, it has algorithms that attempt to eliminate collisions in any circumstances. That's a fairly different approach, which ends up being much heavier in terms of effort and investment.

Make the AI play grand theft auto for many thousands of hours.

Reply on news.ycombinator.com