chiasm

search this place

 

e-mail: j.f.atkinson -at- gmail.com who is this

GO GO

« H E email
posted Wed, 02-09-05



Pejman Yousefzadeh loves GO, linking yesterday to several blogs and articles discussing the game's famous complexity, including this fascinating piece in the Economist on the game's past and very-much present history:

The heavyweight pros on late-night cable television boast nicknames such as Monster, Razor, Butcher, Assassin and Knitting Needle. The most famed matches in history include the Blood Vomiting Game of 1835, the Famous Killing Game of 1926 and the Atomic Bomb Game of 1945. No, this is not some bone-crushing contact sport. It is a simple parlour game where two opponents, comfortably seated and often equipped with nothing more than folding paper fans and cigarettes, take turns placing little stones, some black, some white, on a flat wooden grid. Simple regarding rules and gear, that is, yet so challenging that in this mind-game, unlike chess, and despite the long-standing offer of a $1.6m reward for a winning program, no computer has yet been able to outwit a clever ten-year-old.

I haven't played since the olde school dayes and don't remember it much, and certainly wasn't aware of the immense challenge GO poses for artificial intelligence folks:

"Computer Go is one of the great AI challenges today. Starting from the 1950s it was chess, but after Kasparov lost against IBM's Deep Blue computer, a lot of people lost interest in it," said Graepel.

"However, current computer Go programs can hardly beat even mediocre amateur players," explained Graepel...

To devise a computer program that will play chess against a human, you define a function that tells you how good a given position is for one player. Then you specify all the positions that can be reached with one move and all the positions that can be reached with a given response of the opponent. The computer crunches through millions of positions and evaluates all of them to pick the best move.

That brute force method doesn't work in Go, because there are about 200-300 possible moves you can make when it's your turn. In the first fourteen moves in a Go game, the game tree has 10 thousand million leaves. The program that beat Kasparov could take over one and a half
years to play one move in Go.

"So it's infeasible to enumerate all my possible moves and for every one of them enumerate all the possible moves of my opponent and so forth, because there's just too much complexity," explains Graepel...

"The Japanese have a word, aji. Literally translated, it means "taste."Taste lingers, and so does the influence of a Go stone, even if it appears weak or dead, because of the influence it can have in the future," said Graepel. "It is this lingering and subtle effect of stones that creates uncertainty and makes probability theory an appropriate language for talking about Go."

Modeling the game of Go is akin to modeling one of humanity's most puzzling mysteries - what does the future hold?

However, I have most certainly read about the challenges GO poses for the US defense establishment, in chapter 6 (PDF!) of RAND's essential/seminal/&c 1996 report "The Advent of Netwar" (I just started going through the 2001 sequel, "Networks and Netwar," today, both are free and more or less essential for structuring security thinking 21st century-style). Essentially, old-fashioned and 'modern' modes of maneuver warfare generally resemble chess, with well-defined lines of battle, massed forces, and a strict hierarchy of pieces; 21st century netwars (postmodern war, if you like) are like GO, which is

more about distributing one's pieces than about massing them. It is more about proactive insertion and presence than about maneuver. It is more about deciding where to stand than whether to advance or retreat. It is more about developing web-like links among nearby stationary pieces than about moving specialized pieces in combined operations. It is more about creating networks of pieces than about protecting hierarchies of pieces. It is more about fighting to secure territories than about fighting to the death of one's pieces. It is also less linear than chess.

Annoyingly, for some reason Acrobat won't let me cut and paste text, so I won't bother typing more, but it's just as well b/c the material is too coherent and important to excerpt. Interestingly, earlier in the chapter the analogy is extended to great military theorists - basically, 'less Clausewitz, more Sun Tzu'. All worth your time to read, especially if you've got a printer (or can take advantage of your office printer, in my case) to make yourself a hard copy - it's a huge pain to read as a PDF.


ok, tell me what you *really* think