John Chamberlain
Developer Diary
 Developer Diary · You Heard It Here First · Tuesday 23 March 2004
How Bad are Computers at Igo
Just how bad are computers at the game of igo? Pretty bad. The game of go, or igo as it is known in Japanese, is right up there with speech recognition, autonomous four-runners, and olympic swimming when it comes to things computers are bad at. Top players typically give the best computer igo programs a handicaps in the 20s. A twenty-one stone handicap is sort of like playing somebody in chess with no queen and no rooks. Here is one typical game at an amazing 29-stone handicap. At this handicap even a beginner would have a chance against the best players in the world, but the computer loses anyway. At least it didn't break any rules.

Since there are so many possible moves direct calculation is not possible, at least with today's technology. Therefore the computers have to use heuristics and pattern recognition like humans, but in practice it seems to be impossible so far for programmers to do this effectively. When a human plays igo they tend to think like this: "that group is safe, but my other group needs to connect out or make an eye, if I play move x I can setup a cut that will leave my opponent's group floating, my cutting stones will float too, but I will have the advantage in the fight... etc". In other words there is a knowledge representation problem at work in go that is mixed with calculation (which is called "reading" in the game). One problem is that computers can't read because there are so many possible moves and the outcomes may affect widely different parts of the board. Humans can aim towards particular shapes that they know to be good, but computers just can't put it together.

As the computer chess battle closes in on victory for computers attention now seems to be turning to igo as the next benchmark for strategic programs. It could be a long time before they make headway in this department though.

return to John Chamberlain's home · diary index
Developer Diary · about · · bio · Revised 23 March 2004 · Pure Content