A new computer Go era


A new computer Go era ?

Back to forum

Thibault de Vassal    (2007-03-12)
A new computer Go era ?

It seems a new computer Go era just started...


Quote : "The 19 by 19 board which top players use is still hard for a machine, but the new method is promising because it makes better use of the growing power of computers than earlier Go software."

Quite strange to read about the growing power of computers regarding Go... I suppose programs have much to learn first.. We will see :)

James McKenna    (2007-03-24 13:00:07)
A new era

will a computer ever beat the worlds top players as in chess?

Don Groves    (2007-03-25 03:10:20)
A new computer Go era?

Eventually computers will be able to play Go on a 19x19 board as well as they can now play on a 9x9 board. It's just a matter of spending the time and money to build a powerful enough computer. But, there's no reason why the Go board cannot grow larger. A larger board, say 23x23, wouldn't change the game much for humans but would astronomically increase the time required for expert computer play. But again, if someone wants to spend the time and money, computers will eventually have the power to be the best Go players, just as in Chess.

Thibault de Vassal    (2007-03-25 04:11:00)
Computer Go

I don't agree with that anymore, unfortunately. The more I play, the more I think Go can be learnt by computers, probably at a 2 or 3 dan level.. maybe more. Of course the speed of processors won't change it much. In my opinion programs playing Go will divide (some probably do it already) better their "thoughts" and the game in several parts, some of recognition [shapes, liberties..] and others of calculation [fights that are often reduced to a ~6x6 calculation, whatever the board size] to evaluate positions... So, the board size might be not so important in future and I don't think it will change their level much...

Don Groves    (2007-03-25 05:50:08)
Computer Go

I was thinking about that very thing, Thibault, the idea of using neural networks to analyze the board first in small areas and then higher level networks would combine the small areas into larger and larger areas until the whole board was analyzed. The problem with this approach seems to me that the size of the small areas is variable depending on the depth of the game. Developing algorithms to deal with this variability could be a very hard problem, whereas humans deal with this problem easily.

Lionel Vidal    (2007-03-25 11:24:50)
Computer Go

While I agree that programming Go is much more a problem of algorithm than a hardware one, I think you underestimate the theorical difficulties.

First, a word on the alluded new approach (BTW the french edition of 'Pour la Science' has an article on this algorithm this month, but not very involved): it seems promising only because that program regularly beats other program using what we can call a traditional approach: tree exploration combined with pattern recognition and some clever splitted evaluation function. That is fine, but does not mean much for human, considering the poor level of all these programs.
AFAIY the very best program is said to be at low pro-dan level on a 9x9 (without any concrete real test match, that is with money at stakes... but let's suppose it is true). The problem to play on 19x19 is that the nature of the game dramatically changes: in short the tactics is more complicated and the once very basic strategy of 9x9 becomes overwhelming! There is still no known algorithm to tackle that problem. Such algorithm could exist of course, but don't hold your breath :-)
Now I am quite eager to read the tests and pubications on these researches :-)

The neural network approach is interresting but is more or less stalling (again AFAIY) in recent programs mainly because of a fundamental flaw: the tuning of the gap functions. In Backgammon, where this approach works very well, these functions are tuned by simulation: basically, the program plays many, many games against himself and in a way learns (that is tunes its network) depending on the results. As you may guess, this can not work in Go because of the complexity of the branch tree. So the problem is how to tune the network (and 'by hand' cannot be a soution, believe me, considering the number of nodes and the type of the functions being commonly used!)

Of course I simplified a lot and the maths behind these kind of algorithm are involved enough (and very interresting :-)) that someone may find new ideas that will revive one path or another. But my feeling is that the pros of go have nothing to fear for a long time...
You have to consider that the very best programs are not beaten, but crushed, by multi-dans amateurs, you know, the kind of player a top pro will beat at 5 stones while blitzing and at 9 stones if some money is at stakes :-)

Now I may be wrong, and I remember in the 80s many people saying the same thing for chess, and betting on the fact that a program could never beat a good player in at least 50 years :-)... but at that time, I did not agree :-)) mainly because the algorithms were more or less basically known already... the 80s hardware was a problem, but a technical and not theorical one...

Sorry for that too long reply... I can't believe I typed so much... that must be my new keyboard, and the fond memories of some past jobs ... :-)

Thibault de Vassal    (2007-03-25 18:39:15)
Computer Go

Neural networks will be at most an "extra" or a small part of the solution for a good playing Go program IMO. I don't believe much in it until an artificial brain have the "power" of a human one and we're very far from it. But as in chess (and Rybka's coming), a lot of knowledge probably has to be implemented yet and algorithms to be improved before that. That's why I said it could probably reach a 2 or 3 dan level, which still looks a reasonable level (ok, only an assumption) :) ... Then, it should be much.. much harder.

Don Groves    (2007-03-26 02:04:20)
Computer Go

Yes, we were saying the same things about Chess 30 years ago. This is one reason I feel as I do, the fact that computer capabilities are still increasing just as fast as they did over that period. From experience, using the past 30 years as a guide, I just cannot bring myself to doubt there is much they won't do in the next 30 years. The reason I mentioned neural networks is that they are very good at pattern recognition which must be a large part of the solution if computers are to excel at Go. I do not think a brute force approach, or even a statistical one, will ever succeed at computer Go except in the endgame where they can dominate as they do in Chess once the number of possible moves becomes small enough. But if neural nets can be trained to recognize good shape and then apply basic fundamentals in the opening and early middle game, we already know they will excel in the endgame. It's the latter part of the middle game where I think the action will be in designing computer Go programs -- where fundamentals are not enough and there are still too many moves available to use brute force. But, if computers are ever good enough to get to a winnable endgame, watch out ;-)

Elmer Valderrama    (2007-03-26 11:29:35)
How Life imitates Go

..any takers?

From what you guys write it looks more promising than Life imitating Chess.. --for if Chess can be programmed, can Life ever be ? ;-)

Thibault de Vassal    (2007-03-26 14:58:25)
How Life Imitates Chess

Did you read this book by Kasparov, Elmer ? .. What do you think about ?

Elmer Valderrama    (2007-03-26 18:15:28)
How Life Imitates Chess

Has it been already published?

To start with, something tells me that the title should have been: How Chess Imitates Life, as it is written by a chess player not a philosopher or politician.

Not just because there is nothing bigger than life, but because we would be in real trouble if we had to make use of chess methodology to find out how to make the right decisions in life.

It can help, true, but no more than it did for Napoleon, for example ;)
(--Wellington would be turning in his grove trying to claim a Master Norm for Waterloo ;)

I have real difficulties trying to grasp the link between chess and politics.

Was Churchill a chess Master? if that is so, then Bush and Blair must be ELO 1200 ;-), and Garry must be declared Russian President ipso facto

Which French presidential candidate from your list plays chess?

Which one is considering learning some chess strategy?

Would they get more votes if they declare this intention? -sort of getting into their rights minds and improving their decision making? Hey, Garry is missing some prospective customers here..8-)

Thibault de Vassal    (2007-03-26 18:59:43)
How Politics Imitates Chess

Garry Kasparov is probably not a philosopher, however it is not obvious to me that Chess Imitates Life, as (you said it) there's nothing bigger than life... Chess is a part of life. On the contrary, we could say life imitates chess because some decisions, sometimes, can be reduced (as far as possible) to chess strategies like reality can be reduced to science. The same, How Reality Imitates Science makes sense to me, whereas How Science Imitates Reality doesn't.

"How politics finally does not imitate Chess" by Garry Kasparov should be much more interesting :-) .. with a preface by Vladimir Putin : "How life doesn't imitate politics" :>

In France, the election sometimes makes me think to a.. more than chess, a Go game... I think Nicolas Sarkozy uses some chess(Go)-like strategies and knows openings/Joseki & tactics best. At least he may know very well the work of Arthur Schopenhauer : "The world as will and representation" & "L'art d'avoir toujours raison". In comparison, Ségolène Royal and other candidates seem to use faith and "religion". Anyway, the result should be quite the same as life doesn't imitate politics much nowadays :/

Elmer Valderrama    (2007-03-26 19:18:28)
imitate = model

..nice description of the candidates' campaign strategy, Thibault :-)

How Science model Reality, and How Chess model Life, does it make sense to you? Hope so, since the object to be imitated is Reality or Life, and the imitator must be smaller/lesser by definition. (other synonyms are: imitate = try to be similar to = tend to look like = model)

Elmer Valderrama    (2007-03-26 20:12:01)
The perfect title

As the title couldn't be more awefull, let's try to improve it:

"Decision Making: Invaluable Lessons Taken from Chess"
by G Kasparov

Instant best-seller.

Although the decisions of splitting from FIDE, creating a ghost GM association, not defending the WC in a proper Cycle, and....etc,etc..premature retirement, ..etc ..etc, don't speak too well about the lessons, let's try and improve it:

"Decision Making: Invaluable Lessons Taken from Chess, With Emphasis in What Not to Do"
by G Kasparov

A killer.

Thibault de Vassal    (2007-03-26 20:32:13)

Premature retirement... I don't know but who else than Garry can say ? .. Even if politics is a bad personal choice, it is a respectable one anyway.

Don Groves    (2007-03-26 22:13:24)
Chess and politics

Elmer wrote: "Was Churchill a chess Master? if that is so, then Bush and Blair must be ELO 1200 ;-)" LOL! I can't speak about Blair but you are overrating Bush by several hundred points.

Thibault de Vassal    (2007-03-27 01:41:18)
Chess and politics

Hmm, you sure ? .. George W. Bush may have won another war while losing in Irak. His chessboard may be not yours :/

Don Groves    (2007-03-27 02:55:47)
Chess and politics

It probably isn't known much outside the US but Bush is merely a figurehead, not the brains behind what is going on. Cheney and his gang of neoconservative fascists are the real power behind the throne. But, you are right in the sense that they have done so much damage to US democracy in just six years :-(

Elmer Valderrama    (2007-03-27 17:54:12)
Chess and politics

So Bush's national ELO is even lower ;-)

I assumed they may have won some ELO points by kicking Talibans' ass, but lost a lot more for playing dubious lines elsewhere in an otherwise easily won game --and now the opponent is playing the infamous Dead Man defense in the literal sense :((
Very sad and bad.

Don Groves    (2007-03-28 01:24:06)
Chess and politics

Now the Taliban has regrouped and is back in the game. Bush doesn't know how to finish a won game. Does he really want to win, or does he just want the game to go on forever, keeping his gang in power in the US and making even more money for his corporate buddies?

Thibault de Vassal    (2007-03-28 04:33:21)
Chess and politics

Obviously nobody else want to play with Bush anymore :)

China is too strong already and others play without any oil stakes or prefer to play other games. Poor boy...

Don Groves    (2007-03-28 08:46:47)
Chess and politics

Then other country's leaders must stop cooperating with Bush! Unless they do, the US will continue on this path of world domination. There was a report today that Russian intelligence says Bush will bomb Iran on April 6th. The entire world must condemn this action starting now! US military is overxtended and if other countries will just refuse to support more US aggression, it will have to stop at Iraq and Afghanistan.

Thibault de Vassal    (2007-03-28 08:48:02)
Chess and politics

April 6th ? .. suspense will not be long much.

Elmer Valderrama    (2007-03-29 13:00:51)
Chess and politics

Look no further. Here it is.

The (Chess)Recipe for Success in Bussines and/or Politics
--(fortunately it had already been discovered)--

* Politics is a fairy tale of 1001 bad decisions.

* Bad decisions are all there, waiting to be made.

* Some part of a bad decision is always correct.

* The winner is the one who has made the next-to-last bad decision.

* Losses only prove that someone has erred in his decision.

* Moral victories do not count.

* To avoid losing a man, many a general has lost the war.

* It is always better to sacrifice your opponent's men.

* A threat is more powerful than its execution [the most violated rule]

* There is really only one big mistake - underestimating your opponent [The next-most violated rule]