Hacker News Reader

Seems like the perfect problem to apply reinforcement learning to. Just let it play and learn at the same time.

I think they could even skip the explicit implementation of ball position/velocity tracking.

I think it would be painfully slow to have it learn in real time, especially if a grad student had to put the ball in between every game like what is demoed.

You could definitely pipe (literally) the ball to the input like in real pinball machines and trigger a new iteration.

Reply on news.ycombinator.com