The Electronic Holy War: Computer Go in the New Yorker

[gogameguru]

I found out about this because they got in touch with us for the article. I thought some people here might be interested too.

http://www.newyorker.com/online/blogs/e ... y-war.html

[EdLee]

Hi David, Thank you.

Yesterday I chanced upon this NPR piece, which has no connection to your article except for the part
of the New Yorker. But I liked it very much and will buy the book --

How About Never -- Is Never Good for You ?

[John Fairbairn]

You could tell the writer wasn't a proper go player. He thought four stones was a small handicap. (Ishida lost by 5 points)

Although the march of the machines is unstoppable, I found it interesting that the pros who are now specifically studying 9x9 in response to the 9x9 programs appear to have improved their results (see the series in Gekkan Go World).

There is, I think, another important point these articles tend to ignore (although the one above hints at it), and that is value to human learners. Chess programs think differently from humans but there is enough overlap, especially on the tactical side, that programs can "explain" their moves to a degree that humans find useful. In go, there is so far no overlap at all and so, apart from satisfying the initial curiosity, playing a computer is less fun than playing thwack-a-mole.

Come to think of it, though, next to nothing is as satisfying as thwack-a-mole.

[Fedya]

Hi David, Thank you.

Yesterday I chanced upon this NPR piece, which has no connection to your article except for the part of the New Yorker. But I liked it very much and will buy the book --

How About Never -- Is Never Good for You ?

The pig should say, "My wife is a slut."

[RBerenguel]

You could tell the writer wasn't a proper go player. He thought four stones was a small handicap. (Ishida lost by 5 points)

Although the march of the machines is unstoppable, I found it interesting that the pros who are now specifically studying 9x9 in response to the 9x9 programs appear to have improved their results (see the series in Gekkan Go World).

There is, I think, another important point these articles tend to ignore (although the one above hints at it), and that is value to human learners. Chess programs think differently from humans but there is enough overlap, especially on the tactical side, that programs can "explain" their moves to a degree that humans find useful. In go, there is so far no overlap at all and so, apart from satisfying the initial curiosity, playing a computer is less fun than playing thwack-a-mole.

Come to think of it, though, next to nothing is as satisfying as thwack-a-mole.

John, I guess you'll know the answer (and fits the subject nicely here!): why is there so little about pro-pro 9x9 play available?

[Mike Novack]

I think there may be confusion about what computer programs might be able to explain or not and this is despite a MCTS evaluator being used.

What is being confused are the questions "why is this move a good move?" << what are the go reasons behind it >> and "why is this move (which has go reasons for considering it a good move) better than that move (also has go reasons to make it a good move).

The MCTS evaluator is deciding on which of the moves it is considering is the best on the global game consideration "which is most likely to win the game for me". Look at how a pro considers moves (reading a book like "The Go Consultants" might help). All the moves they are considering are good moves and you can see why. But the decisions "which is best" more intuitive.

OK, a program which uses a MCTS algorithm for its evaluator might be "try everything" (so no go reason behind which moves are placed in the set being evaluated for best) but it might also be using an AI operating on principles of go to narrow down that set. In the latter case the program would be able to explain "what are the go reasons behind the move that was chosen" even though it couldn't say why those reasons outweighed the go reasons behind the other potential moves. Remember, that isn't something that depends on which is locally best or locally more important.

[John Fairbairn]

John, I guess you'll know the answer (and fits the subject nicely here!): why is there so little about pro-pro 9x9 play available?

Wrong question, I think. There's been plenty over the years in Kansai magazines.

But would you expect F1 drivers to be much interested in dodgem cars?

[hyperpape]

Mike, you keep making this point, and it's an ok point, but I don't actually see any evidence that what you're saying applies to anything anyone said in this thread.

[Dante31]

The artical feels a bit sensationalist.

[Uberdude]

Mike, you keep making this point, and it's an ok point, but I don't actually see any evidence that what you're saying applies to anything anyone said in this thread.

Never miss an opportunity to pontificate!

[RBerenguel]

John, I guess you'll know the answer (and fits the subject nicely here!): why is there so little about pro-pro 9x9 play available?

Wrong question, I think. There's been plenty over the years in Kansai magazines.

But would you expect F1 drivers to be much interested in dodgem cars?

Well, seen from afar it seems like there is no pro 9x9 play.

Re: cars, some would be. It's just a close enough field.

[emeraldemon]

GnuGo actually has a fun feature where you can feed it an sgf and it will evaluate every move and tell if it thinks another move is better.

You had to take it with a huge grain of salt, I think I remember someone running it on a pro game and it only thought 25% of the moves were "right", but it was a cool idea. I wonder if someone has thought of adding the feature to a newer engine.

To expand, it might even post longer sequences in situations where it thought you were "more wrong", like offering a refutation almost. It could try to point out your "biggest mistake" or "game-losing move". For 9x9 where machines are very strong, it might actually be pretty useful.

[RBerenguel]

GnuGo actually has a fun feature where you can feed it an sgf and it will evaluate every move and tell if it thinks another move is better.

You had to take it with a huge grain of salt, I think I remember someone running it on a pro game and it only thought 25% of the moves were "right", but it was a cool idea. I wonder if someone has thought of adding the feature to a newer engine.

To expand, it might even post longer sequences in situations where it thought you were "more wrong", like offering a refutation almost. It could try to point out your "biggest mistake" or "game-losing move". For 9x9 where machines are very strong, it might actually be pretty useful.

There are a set of scripts bundled with Fuego or Pachi (don't remember which) which allow you to do more or less this (I had to tweak it IIRC.) It's incredibly slow, though.

[moyoaji]

The article feels a bit sensationalist.

Of course it is. What layperson would want to read this article:

"A go program beat a professional player! However, the handicap was fairly substantial and computers are still years away from being able to take on the best humans in an even game."

[Bantari]

playing a computer is less fun than playing thwack-a-mole.

I was always curious about such statements, both in chess and in go.
To me, when you play an opponent that can beat you, give you good game, and make you sweat to stay afloat - in other words, one who makes very good moves, what does it matter if it is a computer and a human? Sure, there is a difference in how computer evaluates positions and chooses moves, but such differences exist between human players as well. I would even go as far as to say that there is a lot of value precisely because of that difference.

I mostly hear this kind of stuff from chess, in go it is a relatively new "issue". And while people actively look to watch and play strong player (who us us would pass a chance at a game with a pro?) - they seem to look with disdain when they learn it is a computer who makes such good moves and plays strong enough to beat the human players. Suddenly, such games become uninteresting and such opponents undesirable. Why? Good moves are good moves, and if they are better than your moves, you can certainly learn from them, regardless who makes them, no?

Whack-a-mole is for stress relief only, so no real comparison here that I can see.