Slashdot Mirror
Brain vs. Computer: Place Your Bets
dev_null_ziggy writes: "CNN reports that the current chess guru is going up against a supercomputer, amusingly titled 'Deep Fritz.' The match is scheduled for October, and the current champion, Vladimir Kramnik, stands to win $1 Million dollars if he wins. Of course, since he'll be snagging $800k for a draw, and $600k for a loss ... I'll give two to one odds on the machine."
This game clearly shows how stupid computers really are. For your amusement:
White: L. Van Wely, Black: Fritz SSS; played in Rotterdamn 2000
1.c4 e5 2.g3 Nf6 3.Bg2 Nc6 4.Nc3 Bb4 5.a3 Bxc3 6.bxc3 0-0 7.e4 a6 8.a4 d6 9.d3 Bg4?! 10.f3 Bd7 11.Ne2 Qc8?! 12.h3 b6 13.f4 Be6? 14.f5 Bd7 15.g4 Ne8 16.Ng3 Qd8 17.g5 Bc8 18.h4 f6 19.Qh5 Na5 20.Ra3 Qe7 21.Nf1! Nc6 22.Ne3 Qd7 23.g6 h6 24.Ng4 Ra7 25.Rg1! 1-0
>Even more important is the fact that we need not
>search the full search tree (indeed Deep Blue did
>not, using instead something called singular
>extensions).
Deep Blue did not search the full tree, but
singular extensions are a different beast.
Singular extensions let the computer search
_more_ than would be needed.
The idea was to detect horizon effects and avoid
them. The overhead for doing this is large, but
the DB team believed they had so much computing
power anyway that it was worth the tradeoff.
This tradeoff was made in some other places
as well, for example Deep Blue did not use
nullmove pruning, something which nearly
every program nowaways does as which can
prune away large parts of the tree relatively
safely.DB's team decided it wasn't worth the risk
with the computing power they had.
Deep Fritz uses it very aggressively and hence
can sometimes see just as far Deep Blue could,
but also makes more mistakes because of it.
PS. Aske Plaat's proposed improvements are not
used in any top program noawadays. They cause
troule with some of the other tricks in use and
the gain is not large enough to live with them.
--
GCP
>One of the tricks of 'Deep Blue' was a library
>with every game of chess played at the master
>level in the last century. That's what made it
>play like a human.
And Kasparov simply sidestepped this by making
some seldomly played moves at the start. You
can see it easily by looking at the games. The
machines opening play was all but human.
>Kasparov lost the first game because of an error
>in his training, he prepared himself to play with
>a machine and got an almost human player.
It was still a machine, but just with a lot more
chessknowledge and tactical speed than anything
else at that time. He was expecting something
like Fritz (literally!) and got something much
more powerfull.
--
GCP
The biggest advantage of the machine in this kind of games is that it's more difficult for it to make a mistake. I don't know what is the depth of moves that the machine can calculate, but someone at the level Kramnik can usually "see" 10 moves ahead. Then an error screws up everuthing. How long until we get a computer capable of doing this kind of search ? Then we could really see a computer playing a game completly different from a human, and winning ?
It's interesting that the programmer of Deep
:)
Fritz (Franz Morsch) has been mouthing off that
his program is ready for Kramnik and should be
equal to Deep Blue.
They played in the Dutch Championships last year
and couldn't even manage to win. Now they're
saying they stand a chance vs the World Champion?
Well, if he goes too hard on vodka maybe.
This match is simply marketing. They know their
computer is going to lose, but unlike IBM, those
guys actually _sell_ their chesscomputers. And
many people are going to want the one that was
good enough to play the World Champion.
They even 'fixed' the qualifier for this event
so that only their programs played (Deep Fritz
and Deep Junior are both from the German ChessBase
company), nicely blocking out the computer World
Champion (Shredder), as well as blocking out most
other strong contenders (Crafty, Tiger, Rebel,
Hiarcs, Nimzo, Diep, etc...) on false grounds.
So, please don't say this match is anything like
Deep Blue - Kasparov. Fritz is significantly slower
and stupider, no matter what they would want you
to believe. This is in no way the best chess
computer to have ever existed.
Also, don't say this is the end of human
intelligence
if Kramnik loses. Not until a go program starts
beating me, at last
--
GCP
Under the F.I.D.E. laws (I dunno how official this is, since I'm not a chess person -- it seems to be "official" chess, according to the site), rule 10.10 states that it's a draw when the chess board repeats its state for the 3rd time. There are OBVIOUSLY a FINITE number of chess board states (placing a finite number of pieces on a finite number of squares, plus a few extra bits to represent piece "rights" such as castling and en passant stuff). Therefore, sooner or later, a chess game will either end "normally" or run out of states that haven't been hit twice.
My initial reaction to this is that the machine would give the human grandmaster a sound thrashing. But the more I think about it, the more I wonder. Negating years of experience for the GM would be a blow, certainly. But it would depend on how the rules were changed. 1. If you changed the way pieces move I think the GM would be lost. To the machine such a change is merely calculation, to the GM its an entirely different game. 2. If you changed the game so that the average branching factor was huge the machine would be lost. In a chess middlegame the branching factor is around 35 moves... hence fast microcomputer engines can see from 10-15 ply deep. But computers suck at go where the branching factor is enormous -- an alpha-beta tree search is hard pressed to get very deep when the number of positions explodes more quickly. One thing that many chess servers have is called "Fischer Random" chess where the pieces on the back rank in the starting position are scrambled. This removes fixed opening lines but keeps the pieces moving the same way and the same maxims of chess. BTW, my money is on Kramnick. "Deep Fritz" is a 5 million position per second beast on a parallel x86 machine. Deep Blue was a 200+ million position per second monster on custom hardware. How is fritz the "successor" to deep blue?
Other languages
Not quite true. Master players are much faster and more acurate analysing random positions too. See here for example
I don't know where you got the statement that
Deep Fritz beat Deep Blue, but it's obviously
false given that Deep Blue never played anyone
but Kasparov.
There were single-chip versions of Deep Blue on
the web for a while, so it could be that they beat
such one. But its more than 400 times slower than
the full Deep Blue.
Also, the win vs. Kasparov was in a blitz game.
Computers have long been superior in those fast
games.
This is marketing people. Many here don't seem
to realize chess in multimillion business, and
lying is ok if it makes you sell better.
--
GCP
>and they rule at Connect 4
This game has been SOLVED by Victor L. Allis.
He also invented a new tree search algorithm
which is extremely strong _when_ it can be
used.
He used a combination of this tree search and
rules (black can't win if this parttern
is present etc..) to solve it.
>about 30 at chess, 10 at checkers,
It's 38 for chess, 2.7 for 8x8 checkers (where
a comp is already world champion)
The use of tree search depends on a lot on
the tactical nature of the game. You can still
use it with a branching factor of over 100 if
the game is tactical enough. (so 5-7 ply searches
beat most humans)
But go needs more longtime planning, and you need
way more depth for that.
--
GCP
A sufficiently powerful computer will always beat a human opponent, but creativity is important for the human if he is to have a chance.
In the early days, say early 80s when the computers took their first steps in being proper opponent for good chess players humans those computers one by one by useing their lack of brute force and/or intuition against them. Boris Spasski whooped one computer beautyfully by sacrificing few king side pawns at point where even a moderate human chess player would've realized that by giving room to Spasski's rocks there'd be problems _in the horizon_. The opposing computer those days naturally couldn't predict that and Spasski indeed launched a glorious attack and won.
That was just a good example how humans usually play against computers. And this is also what Gasparov tried against Deep Blue but in vain. A nice example of where computers had gotten at that point was in one of the games, where Kasparov launched a really promising attack on the king side. It really was promising at that point and most likely any chess guru who was capable of spotting that offense opportunity would have seized it. BUT, at the decicive moment when deep blue had to decide wether to fall back and just try to minimize the damages or call the bluff it (DB) had calculated _every_ possible ending that attack could result in (and we're not talking about checking mates in 5 or 6, but serious amount of prediction) using nothing but brute force. Thus the Deeb Blue took the pawn Kasparov had sacraficed and dealth with the attack to a point where Kasparov gave up.
The throne of chess has been lost for good. There's little reason to suspect that Deep Fritz would loose unless it's significantly slower (or it runs M$ software) than Deep Blue. Garri Kasparov was by far good enough to represent our kind...
(every little detail in this comment may not be 100% accurate as I can't be arsed to check the references right now, but it's by far close enough)
-
1 Earth is warming, 2 It's us, 3 it's royally bad, 4 we need to take action NOW