[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [computer-go] Modern brute force search
At 19:33 7-11-2004 -0500, Don Dailey wrote:
>
>You brought up an interesting point.
>I used to speculate with Larry Kaufman about how to program an infinite
>speed computer to play chess.
>The obvious way would be to search to the end of the game and return
>at leaf nodes -1, 0 or 1. This would produce an "omniscient" player
>which would never lose a game even to "god."
>
>But would that be the best way to play against a human? Perhaps the
>game theoretic value of chess is draw, which many believe is likely.
>The omniscient player might immediately trade down all material to
>obtain a draw very quickly, even before the human had a chance to
>screw up. Or, it might offer a draw on the first move since a draw
>offer is in the set of best possible moves. Game programs usually
>assume best play by the opponent, which is the only correct way to
>play.
>
>But it would be easy to imagine an omnicient computer applying
>heuristics which minimize the opponents chance of drawing, for
>instance by avoid simplification, mixing things up, keeping the game
>as open as possible etc.
>
>I admit I'm not an expert on GO by any stretch, but I have been
>speculating that an "omniscient" computer GO player would be far ahead
>of human masters. I suspect that an omniscient GO player could be
>programmed to play in the same way strong players beat weaker players
>with handicaps, as you say, "leaving everything unsettled for as long
>as possible."
>
>It sounds like GO isn't as complicated as I thought it was, due to the
>fact that many of you believe masters are pretty close to optimal play.
>
>For instance, Jonathon Schaeffer has taken a game very MUCH simpler
>than even chess, and produced a program that is stronger than the best
>human players. This game is called "checkers" and is played on an 8x8
>board. In the USA just about every child knows this game and it seems
>so simple compared to chess. In his book, "One jump ahead" Jonathon
>noticed that no matter how deep the program looked, it seemed to
>benefit from looking even farther. There was always something it
>needed 2 or 3 more ply to see.
8x8 Checkers is a very simple game and Jonathan a very cool guy.
Checkers is very hard to compare even with other checker forms such as
international checkers.
international checkers there 3 vs 1 queen is a draw.
In checkers such endgames are much simpler unless a queen is on the long
diagonal.
Checkers is all about zugzwang, unlike chess and go. In chess zugzwang only
happens in far endgames some, and even then only when there is not a single
piece left.
So in checkers it's simply over after a few moves and you can very clearly
read it.
A professional checkersplayer such as Tinsley could play the game near
perfect.
Because this game is nearly solved now with databases, you can analyze his
games (in the good days that he wasn't dead sick) and will see that he
really played near perfect.
Checkers is such a simple game that a professional player CAN play perfect.
Of course you need many professional players before a talented professional
player is there that can do it.
I agree with you on go in this sense that the openingsmoves a perfect
playing player (a database), will always win from humans.
Note that the opponent modeling is already for years a well known
publication from Jaap v/d Herik and others. See ICGA yearbooks and journal
(www.icga.org).
Where in the start of the game, go definitely is played real poor by the
professional players when compared to optimal play, short after opening i'm
sure that the strongest go players will not be easy to beat by the 361
stone database.
The ending sequences you *can* read real deeply without much of an effort
from professional viewpoint. Even if it's a 150 ply bestmove. No problem
for a pro to FEEL what is the best move there.
In that sense chess for sure is played less perfect during the game and
more perfect in the openingsmoves.
It is difficult to draw conclusions from that in a go mailing list.
computer Go players might not like the conclusion.
The go professionals however will get a smile on their face which never
will get away when they see a professional chessplayer sweat hard on a game.
Their endgame is just near perfect i bet.
Especially in far endgames when the computer never falls for exchanges
towards 'known' won or drawn EGTB positions there the databases just
mercilous kills even professional chessplayers, or mercilous draws them
when the position was won for the professional player.
Of course this is when analyzing their own games.
There is too little official matches human-computers to have good
datapoints what strength they have at what time.
In general humans just don't care a penny as they just get their paycheck
and leave.
But when i analyze with professional players, and nowadays they LOVE to
analyze with computerchess experts, as they have good hardware and all
electronic stuff at hand, you see the bitter truth there.
It is only the utmost world top grandmasters which really are a level
better in far endgames than the rest of the grandmasters.
I really found that out when analyzing recently with a friend of mine.
Only world top really can deal with that.
>It seems this trivial game was deeper and more profound than anyone
>believed.
>
>I see chess as being much deeper and much more complicated. I see GO
>as being much more profound than chess. Am I wrong about this?
>
>If it's the case that GO is simpler than even Chess, I would retract
>my speculation on this. I know that it is often held that GO is
>simpler than chess for humans, just harder for computers. But I don't
>agree with this. A strong chess player who is also a realtively good
>GO player and an expert at many games (Larry Kaufman) once told me
>that there are many more levels of skill in GO than in chess. In
>other words, as he explained, the median GO play has less of a chance
>beating a top GO player than a median chess player has of beating the
>top chess player.
>
>I have been told that it takes longer to "master" GO than chess,
>although I don't really know what this means. I know it's possible
>for a very young person to be world chess champion. How old are the
>world champion GO players?
>
>So my view (and I'm sticking with it) is that we may someday find out
>that great GO players are not as great as we think they are, but it's
>just my opinion. I certainly don't believe they are playing anywhere
>near optimal, although it probably seems that way to us mere mortals.
>
>- Don
>
>
>
>
>
>
>
> X-Sender: rtayek@xxxxxxxxxxxxxxxxx
> Date: Sun, 07 Nov 2004 15:42:21 -0800
> From: Ray Tayek <rtayek@xxxxxxxxxxxxxxxxx>
> X-Scanned-By: MIMEDefang 2.42
>
> At 10:51 AM 11/7/04, you wrote:
> >...
> >And that is a shame. I strongly feel that even when humans are clearly
> >surpassed, we are just getting started. Chess is an incredibly deep
game
> >and the very best humans are very weak compared to the "ultimate player."
> >
> >Just my thoughts. All of this is even more true of Go. There exists
> >(in principle) a future GO player than can make the very best Go
> >player (of today or yesterday) look like a baby.
>
> i don't think so. if by "look like a baby", you mean able to give a large
> handicap (say 6 stones), then no way.
>
> i don't see any evidence that human players can become that much stronger
> than todays professional 9-dan's and i doubt that a program will beat
them
> for a *very* long time (think about "asking" moves, or watch a 9 dan pro
> give an amautuer 7 kyu player a nine stone handicap - the pro will leave
> everything as unsettled as possible for as long as possible).
>
> thanks
>
> ---
> ray tayek <http://tayek.com/>, co-chair <http://www.ocjug.org/>, actively
> seeking telecommuting work.
>
>_______________________________________________
>computer-go mailing list
>computer-go@xxxxxxxxxxxxxxxxx
>http://www.computer-go.org/mailman/listinfo/computer-go/
>
>
_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://www.computer-go.org/mailman/listinfo/computer-go/