Incredible news, I really don't quite believe it... But it looks pretty legit so far.

http://wayt.synology.me/wordpress/1348-2
https://www.youtube.com/watch?v=g-dKXOlsf98
http://deepmind.com/alpha-go.html
https://storage.googleapis.com/deepmind ... ing-go.pdf

AlphaGo played a 5 game match with Fan Hui and won 5-0. 19x19 board, no handicaps(!)

From the video it sounds like their major innovation is to use two neural nets instead of just one. One is a 'policy network' to select move candidates, this is what Facebook did. But then they created a 'value network' that evaluates leaf nodes. Other bots had been using Monte Carlo playouts for their evaluation.

- Kim Myungwan will be doing a review of the AlphaGo vs Fan Hui 2p games on the AGA Youtube this Friday.
- Google DeepMind, the British artificial intelligence company which developed AlphaGo, has issued a $1 million challenge to Lee Sedol from South Korea, the top player in the world for much of the last 10 years, to play in March.

This is huge. I hope somebody with Nature access can give us some details about under what conditions the games were played.

EDIT: Never mind the paper is available for free from Google's website:
http://deepmind.com/alpha-go.html
https://storage.googleapis.com/deepmind ... ing-go.pdf

:O

What????

Ok if it defeats Lee Sedol then

Google is my hero!

https://www.youtube.com/watch?v=SUbqykXVx0A

Another video.

It's an interesting experience seeing videos of Fan Hui 2p and Matuesz Surma 1p playing go on national television. I think that comes rather close to a satisfying verification.

_________________
On Go proverbs:
"A fine Gotation is a diamond in the hand of a dan of wit and a pebble in the hand of a kyu" —Joseph Raux misquoted.

From the paper, the game time was 1 hour main time + 30s byo-yomi. This is no blitz trick.

I'd be curious to know how much they had to pay Lee to participate in this.

_________________
Still officially AGA 5d but I play so irregularly these days that I am probably only 3d or 4d over the board (but hopefully still 5d in terms of knowledge, theory and the ability to contribute).

One stat from someone who worked on the project (https://news.ycombinator.com/item?id=10982008): they can give Crazy Stone a 4 stone handicap and win 80% of the time. I'd still suppose they're an underdog against Lee Sedol, but maybe not a hopeless one?

_________________
Occupy Babel!

I asked Demis Hassabis about processor power, and it might be 200 cores. So, a little while yet.

http://www.bbc.co.uk/news/technology-35420579

Ranked 3rd on top stories, but those ranks change all the time I guess (a lot like my ranks).

_________________
On Go proverbs:
"A fine Gotation is a diamond in the hand of a dan of wit and a pebble in the hand of a kyu" —Joseph Raux misquoted.

Super exciting news, both because of the AI development and also because it may raise some hype and awareness about Go in the US.

I couldn't resist looking up Fan Hui on Remi's goratings: http://www.goratings.org/players/1480.html

He's at 2916 elo, and Lee Sedol is at 3515, so by rating Fan Hui would have about a 3% chance of beating Lee Sedol. This seems about in line with the team jubango Lee played last year, where he beat top westerners at 2 stones. So beating Fan Hui is definitely an accomplishment, but Lee Sedol represents a significant step up in difficulty for AlphaGo.

That said, I doubt anyone predicted we would see a match like this so soon, Go AI is developing incredibly quickly. Is it too much to ask that google would make the AI open source? I'll have to read the paper...

The games.

According to the paper, Fan Hui won two out of 5 informal games played at shorter time controls. If anything, it seems the computer plays better at standard time controls.

4th and 5th

According to the paper, the distributed version of AlphaGo was running on 1202 CPUs and 176 GPUs. Their non-distributed version was running on a machine with 48 CPUs and 8 GPUs. So yeah, this isn't off the shelf hardware. But if CPU power is the only limiting factor, I think we can expect some VERY strong programs available for home computers in the near future.

I just briefly went through all five games. My observations
- Fan Hui's openings are very traditional and peaceful. AI probably knows all these openings very well from past pro games. So there is nothing human player can gain at this stage. It remains to be seem how the AI reacts to the more 'modern' and 'imaginative' ways younger Chinese and Korean players play (for example how would the AI reacts to situations as in this game: http://www.go4go.net/go/games/sgfview/52548 (moves 28 through 68).
- Fan appears to make a number of mid-game mistakes. But so does the AI player (for example, one obvious mistake was move 60 in game 3).

It definitely seems to pull openings from pro games, which makes sense. It played the large avalanche perfectly, and the joseki after microchinese - split - attach that Fan Hui used both times. I also noticed AlphaGo played nirensei both games as black, making a sanrensei for one of them. It's interesting to me because black nirensei isn't that popular with pros today.

AlphaGo does vary its openings at least some, since it faced an identical position and chose the avalanche for one game and the hane for the other.

Would be interesting to see strong pros review this game.

This is probably a high insei level game in asia.

Some quote from the paper
https://storage.googleapis.com/deepmind ... ing-go.pdf







AlphaGo was using 48 CPUs + 8 GPUs. CrazyStone had 32 CPUs, Zen had 8 CPUs, Pachi had 16 CPUs. They were using 5 seconds per move.

I wondered about that. Playing out ko threats is usually poor, especially when they lose endgame points. But "making the next threat bigger" can be justified. (There was a ko top right, and I don't suppose it was over the analysis horizon.)

But I doubt we're going to be able to get inside the head of this type of algorithm, with our usual concepts. Over the five games it appears to be at home in the middlegame.