Keep us posted!

Great News! I still believed computers weren't able to play high profile Go.
Good to hear that programmers are catching up

I don't know much about Go, but what's the deal with "handicap"? That doesn't seem fair when it comes to deciding the winner between man and machine.

Handicap makes sense. This isn't really the ultimate man vs. machine match, its purpose is more to explore just how good a modern program can be on a powerful cluster. I would be very surprised if Mogo won an even game, but I hope we will see a good improvement from the 1k it is at on conventional hardware.

But what the final match will show is whether man underestimates machine. I think we do

8 dan is a pretty serious rank, is it not? How many players in the world are above that level?

Many many players.

That is 8 dan professional. There are many people above that... but not that many. According to http://senseis.xmp.net/?Top20EuropeanPlayers, no one in Europe. It is in any case solidly beyond where the vast majority of go players can hope to reach.

Thanks for the info!

This is super cool. I'm loving these new monte carlo algos, and this is the best human player that I've seen one pitted against.

I'd not say this is a fight of man vs algorithm... We are talking about a 3000 core cluster! With that much processing power, a chess engine would crush a professional like a grape. Nobody is doing it because we know what the result would be, and nobody really likes to see it.

If anything, seeing that an AI based on running MC simulations is the best we can do for go is nothing but disheartening. It's using sheer speed with relatively little reasoning. It doesn't teach us much, other than to invest in faster processors. We'd get more out of a program that learned how to solve Life and Death problems efficiently.

Mogo is hardly just Monte-Carlo. Technically, it's not even UCT, because it doesn't use confidence bounds. Playouts are guided by hand-made patterns, and it apparently has new techniques to handle capturing races, which MC programs are known to be bad at.

Any AI is going to be disappointing once you know what's "under the hood". But isn't playouts closer to how humans operate? We figure out which moves are good by learning basic ideas and theory (vaguely similar to Mogo's patterns) and playing lots of lots of games, which is also what Mogo does. Only we are far better at generalization, therefore we learn far more from far fewer games, and can keep our knowledge across sessions.

So it's easier for a computer to beat a human playing chess than playing Go? Why?

Many more possible games (game-tree), and many more effective strategies and reading required.

Björn, it's a complex issue, but...

1. Go has a huge branching factor. There are just far more legal moves in most positions. Looking at all possible moves, all possible counter-moves, all possible counter moves to that again etc. is just vastly less feasible.

2. Go games last a long time. I don't know what the average number of moves in a game is, but I doubt it's less that 200. That's more than twice as long as chess.

3. There are no (or very few) "instant death" positions in Go, like checkmate in Chess or establishing a connection in Hex. This makes certain powerful search techniques (so-called proof search or sequence finding) less useful. It's not possible to trim the search tree to the same degree.

4. In chess, there are some very solid rules of thumb that you can use to see who's going to win even if you are awful at looking ahead: Material. Less obvious, but also quantifiable: pawn formations. One particularly powerful family of techniques for search, namely alpha-beta searches, is completely dependent on such rules of thumb.

5. It's possible that we humans use our visual systems in clever ways, for shape-recognition etc. so that it is simply an easier game for us than chess. We are well-wired to deal with shapes, we are far less well equipped to do the kind of reasoning chess requires. In other words, it's not just that it's hard for computers, it's easy for us as well.

Why has there been so huge advances recently (from 10k to 1k in maybe two years)? Because a new technique besides alpha-beta and proof search has been discovered: playout-based algorithms. These are able to play a lot of games pretty well without any more information than the rules. They are also more flexible with regards to such rules of thumb as alpha-beta needs - they can use different, softer rules of thumb - which are far more plentiful in Go than hard ones.

8 Dan is very very high rank even in Korea.
There are 17 8dan and 49 9dan players in Korea. 9dan is called "Ibshin"(means entering God skill).

I really don't understand why the 8 dan player do that kind of silly match. Maybe test the program for better upgrade skill i guess.
As far as i know the strongest Baduk program skill is about 6 kyu.(Weakest---- 18.17.16.15......2.1 kyu and then 1 dan, 2dan....9 dan--strongest). That was couple of years ago, maybe there might be better program right now. However not more than 5 kyu.
A lot of elementary students have that much rank(6 kyu)in Korea.

Learnig Go game is like a lifetime study for me.

Surprised to know that computer program skill has been improved to 1K !! What is the program name ?
Eager to see how the program is. So many my Korean relatives complain that computer baduk program is so stupid.

It's official.

Computer > 8 dan

http://www.usgo.org/index.php?%23_id=4602

What does "nine stone game" mean?

Very Surprised to know that!!! 2 or 3 dan!!!

I am also wondering what the 9 stone game is.. Computer puts its 9 stones already on board before the match? It sounds like it.

Anyway, there have been big improvements of Baduk AI!!!

Computer (black) gets 9 handicap stones, placed on the 9 dots on the board.

Not. Not with a 9-stone

"Kim easily won two blitz games with 9 stones and 11 stones and minutes and lost one with 12 stones and 15 minutes by 3.5 points "

The bot's username on KGS was MoGoTitan. Record of the final game can be found here:

http://files.gokgs.com/games/2008/8/7/MyungWan-MoGoTiTan-4.s...

I'm surprised every time I see experienced Go players who haven't heard of the new Monte-Carlo programs You can try them in the Computer Go room at KGS, if you wish to make up your mind about their skill yourself. MoGoBot1 is MoGoTitan's little brother, it has a 2k rank at KGS, established after lots of games. MoGo never sleeps.

That ought to count for something. After all the easiest way to gain rating seems to be to rely on rating drift, I'm 15k already!

Bear in mind their quirks: They will play odd moves when they think they're behind, and they don't bother to secure additional territory if they think they will win anyway (that's why they tend to win with 0.5 points!). They are somewhat worse at handicap Go because of these two issues.

I'm a (very) casual Go player, but can someone explain to me why it's a big deal for a computer to beat a human with 9 stones already on the board? Isn't that an enormous advantage for the computer?