Agronesia

Go is a board game with very simple and elegant rules, but with a mathematical complexity far far above chess. It is a game with lots of pattern recognition involved, so a skilled human player is particularly good at intuitively making judgments just by glancing a position.

We all know how Gary Kasparov was beaten by Deep Blue some time ago, so in essence computers now can play better chess than human. Go, with its enormous computational complexity, is regarded by many as the last bastion to defend human intelligence in strategy games.

Around 5 years ago, the best computer program can still be defeated by a professional even with a 25 stone handicap (e.g., getting free 25 moves at the start of the game). Now it plays much much better, and at the recent U.S. Go Congress (7 August 2008) a supercomputer beat a professional player at 9 stone handicap. Here’s the details of the match:

• The program is MoGo which is based on the Monte Carlo technique.
• It doesn’t use your average home PC, but a borrowed European supercomputer with 800 processors each at 4.7 GHz. The computer’s processing power is 15 teraflops, which is actually very far from the top 10 supercomputers in the world. Still, the machine is a monster by your average Joe’s standard.
• The human player is Kim Myungwan ranked 8-dan professional. The highest ranking for a pro is 9-dan so he is quite strong. However compared to top world players like Lee Changho the difference in skill is far. (There are lots of 9-dan professional players, but even among those we can clearly see few who are definitely champions out of the strongest)
• Kim Myungwan used only 13 minutes thinking time while the program used 55 minutes.
• The human player lost by 1.5 points.

So, it is indeed a progress, but still far from actually getting on par with humans. First of all, the program must get stronger by 9 stone handicaps to play even with human. People say that a handicap stone is roughly equivalent to a 10 point advantage. So the program would most certainly get beaten easily in an 8-stone handicap game by Kim Myungwan. Second, to claim that a computer plays better than human it must defeat the strongest human players. An 8-dan professional player like Kim Myungwan is certainly not weak, but he is far from the top go players in the world that frequents international titles.

The use of Monte Carlo, a soft AI technique involving simulations of random moves, is also interesting. OK, it plays go well. But does it understand go? Or is that question irrelevant, e.g. “If it plays go, then it PLAYS go”.

Tags: AI, Intelligence, Monte Carlo

This entry was posted on Sunday, August 10th, 2008 at 8:11 am by Agro Rachmatullah and is filed under Computer, Go. You can follow any responses to this entry through the RSS 2.0 feed. You can leave a response, or trackback from your own site.