News for hackers. Stuff that matters.

A comment on Alpha Go Zero: How and Why It Works. Parent comment

Re:Alpha Go Zero: How and Why It Works by gwern on Saturday November 04, 2017 @10:46PM (#15627936)
> These are the two biggest distinguishing characteristics of AlphaGo Zero compared to previous AlphaGos, and the OP doesn't discuss either of them.

David Silver disagrees. The most critical distinguishing characteristic is the expert/tree iteration which makes stable self-play possible at all.

[ Parent | on HN ]
Re:Alpha Go Zero: How and Why It Works by cosminro on Sunday November 05, 2017 @03:11PM (#15631303)
A lot of people reading the paper miss this. I guess it's not emphasized enough.

In the first paper, the selfplay trained policy is about 1500 in elo rating, while darkforest2 a supervised trained policy from Facebook is around the same, if not better. So selfplay wasn't of much use the first time around. While in the AlphaZero paper the selfplay trained policy has about 3000 elo rating.

[ Parent | on HN ]
Re:Alpha Go Zero: How and Why It Works by gwern on Sunday November 05, 2017 @03:42PM (#15631437)
> A lot of people reading the paper miss this. I guess it's not emphasized enough.

Yeah, it's hilariously underemphasized. 1 sentence, literally. Fortunately I was able to ask Silver directly and get confirmation that it's the tree iteration: https://www.reddit.com/r/MachineLearning/comments/76xjb5/ama...

[ Parent | on HN ]