Gifu Challenge 2005

Official Web Site: http://www.computer-go.jp/gifu2005/English/

The Nov. 7 issue of the Asahi Shimbun reported on this tournament as follows.

In the shogi world, pros have been prohibited in principle from playing against computers. This is proof of how strong the AI software has become. What about igo? Will programs eventually be able to win against pros, and if so when?

Weak at tough fights, still at amateur shodan-level

Participating in the Gifu Challenge 2005 were 11 Japanese programs, two each from the US and China, one each from North Korea and Holland, and one developed multinationally, for a total of 18.

The result was the third straight win for North Korea's KCC Igo, but Hiroe Nobuyuki 9-dan commented that the top programs have achieved on average the shodan level.

David Fotland's Many Faces of Go, which came in second, played against an middle-school amateur 1-kyu in an open game and won by three-and-a-half points with Black.

Tanaka-kun, the player in question, commented that the software was easy-going and did not really come out fighting.

Fotland is a US 3-dan who has been playing for 30 years and programming go for 23.

He shrugged, "What makes creating a strong program difficult is that there is no fixed strategy. I give my own program nine stones. Whether or not I'll be able to play it even sometime before I die is, uhh, a good question."

The Nihon Kiin assigns ratings to software and currently ranks three programs at the shodan level. Another unranked program is also believed to be shodan. A representative stated "computers are strong against players who play obvious moves, but weak in complicated fights".

25 years from now at the earliest; harder than shogi

Professor Matsubara Jin of the National Hakodate Future University, an AI expert, states that igo software should be able to beat the Meijin by 2030 at the earliest or 2050 at the latest.

He states that "current software is at the amateur shodan level and getting stronger very slowly."

The problem lies in the characteristics of the game. Winning involves gaining the most territory. Shogi has an obvious target of capturing the king, whereas winning at go involves the non-trivial target of grabbing a number of territories.

In addition, the number of positions and moves is much larger than in shogi. Current programs use a shogi-like approach to search. They select the best move from among a number of candidates. In a half-point game, they can corrrectly choose from the limited number of moves. But the breadth of the middle game makes it difficult to select the best move there.

Shogi software, on the other hand, is making great strides. "It should be able to beat Yabu in ten years. In mountain climbing terms, shogi is 80% of the way to the summit, compared to igo which is still at the 10% or 20% level."

The professor has a 5-dan shogi diploma, is about 2-3 kyu at go, and is 1-2 dan at chess.

Comments

John F. Thanks for the report, Bob. I think the name of the chap is Matsubara Hitoshi. I presume I'm not alone in not really believing the shodan mullarkey. As Fotland implies, once you know the weaknesses you can make computers look silly - but how strong is Fotland?

HermanHiddema Fotland is AGA 3D (as mentioned in te text). I think giving 3 (or 4) computer programs as 'amateur shodan strength' is complete bollocks. Plenty of amateur shodans can give the strongest programs a 9 stone handicap and have a good chance of winning. Even taking into account that japanese ranks are not as strong as those from other countries (see RankWorldwideComparison), I do not believe programs have reached amateur shodan strength.

dnerra: Of course 1D is non-sense. The game of MFGO against that 1k was pretty typical for the strong points of computer programs: the 1k played very passively, avoiding all fights, basically making the game one large endgame. Computer programs are a lot better than humans (of comparable strength) at simply picking the largest available move when there are no tactical complications.

RiffRaff: Even without knowing the weaknesses, most Go Playing Programs look pretty silly. I think I'm pretty comparable to most of them, at the low DDK level on KGS. Based on the progress made in the past decade or so, I imagine that the gains made on the computer front will be extremely slow and painful. Especially considering that one of the best potential improvements for many of them is doing pattern matching based on professional moves (i.e. blindly imitating pros and hoping it can actually play the resulting position). I don't really expect to see a true shodan level program within the next 25 years.

Harleqin: Is the game record of that 1k-MFOG game available somewhere?

RiffRaff: Their web site is here and it looks like there is a zip file with SGFs of all the games played. One of the games in it is 'human-mfg.sgf'.

John F. Maybe playing in public was stressful for the 1-kyu human, but I find it hard to rate this game above 12-kyu. The pros have a history of being extremely generous with their ratings for computers (go and shogi), but this seems like a very early Christmas present from Hiroe.