That should be good ......

What do you guys think ?

He will play 6 games...

I'd put my money on Deep Fritz. I had CM9k analyze two of the games from
his match with Topalov, Games 9 and 10. He did not play with great
precision in either game with the settings I used, 4Mb hash, 180 seconds
per move. Compared with what CM says is the best move, Kramnik scored 53%
in Game 9 which he lost and 64% in Game 10 which he won.

If these two games are representative, he may not be playing with
sufficient precision to beat a GM program.

Wilma


(SAT W-7) wrote in news:9136-4529A915-1977@storefull-
3336.bay.webtv.net:

That should be good ......

What do you guys think ?

He will play 6 games...

Wilma wrote:
I'd put my money on Deep Fritz. I had CM9k analyze two of the games from
his match with Topalov, Games 9 and 10. He did not play with great
precision in either game with the settings I used, 4Mb hash, 180 seconds
per move. Compared with what CM says is the best move, Kramnik scored 53%
in Game 9 which he lost and 64% in Game 10 which he won.

If these two games are representative,

Which is a very big *IF* Two games seems to me to be inadequate to draw
any significant conclusions.

he may not be playing with
sufficient precision to beat a GM program.

Wilma

Programs are good at tactics, but less so at strategy. Hence such raw
statistics are not as meaningful as they might appear to be.

Crafty is not the strongest chess program around, but I doubt it would
be that difficult to compute similar statistics for 100's or 1000's of
games with crafty, with no need for someone to laboriously do the work.
Much of it could be automated.
--
Dave (from the UK)

Please note my email address changes periodically to avoid spam.
It is always of the form:
Hitting reply will work for a few months only - later set it manually.

http://witm.sourceforge.net/ (Web based Mathematica front end)

Wilma wrote:
I'd put my money on Deep Fritz. I had CM9k analyze two of the games from
his match with Topalov, Games 9 and 10. He did not play with great
precision in either game with the settings I used, 4Mb hash, 180 seconds
per move. Compared with what CM says is the best move, Kramnik scored 53%
in Game 9 which he lost and 64% in Game 10 which he won.

If these two games are representative, he may not be playing with
sufficient precision to beat a GM program.

Wilma

Only 4 MB hash for analysis? That's nowhere NEAR enough. For 3 minutes
per move (assuming a reasonable fast computer, let's say a 2.4 GHz) you
should have 24 MB of hash to be safe.

Here's how to test how much hash you need. Find any reasonably complex
middle-game position and turn on the Mentor window. Count how many
positions are analyzed in 10 seconds. That is how many BYTES (rounded
upward to the nearest MB) you would want to allocate if you were going
to analyze for 10 seconds per move.

For 180 seconds, simply multiply that amount by 18. Make sense?

jm

Okay. I'll do that for my 1.66 computer. I'm getting a 2.8gh at the end
of October with a 1gb memory, so I should be able to get a more accurate
measure then. Or, I'll just use 64mb hash on both computers. Thanks,
John.

Does it nake any difference that a player may not have the equivalent of
64mb in his head? Probably not, since we're looking at relative
precision between players and need only to be sure I use the same hash
table for all players and games.

Wilma


wrote in
oups.com:


Wilma wrote:
I'd put my money on Deep Fritz. I had CM9k analyze two of the games
from his match with Topalov, Games 9 and 10. He did not play with
great precision in either game with the settings I used, 4Mb hash,
180 seconds per move. Compared with what CM says is the best move,
Kramnik scored 53% in Game 9 which he lost and 64% in Game 10 which
he won.

If these two games are representative, he may not be playing with
sufficient precision to beat a GM program.

Wilma

Only 4 MB hash for analysis? That's nowhere NEAR enough. For 3 minutes
per move (assuming a reasonable fast computer, let's say a 2.4 GHz)
you should have 24 MB of hash to be safe.

Here's how to test how much hash you need. Find any reasonably complex
middle-game position and turn on the Mentor window. Count how many
positions are analyzed in 10 seconds. That is how many BYTES (rounded
upward to the nearest MB) you would want to allocate if you were going
to analyze for 10 seconds per move.

For 180 seconds, simply multiply that amount by 18. Make sense?

jm

"Wilma" schreef in bericht
. 125.201...
I'd put my money on Deep Fritz. I had CM9k analyze two of the games from
his match with Topalov, Games 9 and 10. He did not play with great
precision in either game with the settings I used, 4Mb hash, 180 seconds
per move. Compared with what CM says is the best move, Kramnik scored 53%
in Game 9 which he lost and 64% in Game 10 which he won.

If these two games are representative, he may not be playing with
sufficient precision to beat a GM program.

Wilma

I don't really get the point, because Kramnik doesn't play the same moves as
the pc he plays worse? Maybe he just plays better moves?

Yes. The assumption is that the computer makes more precise and correct
moves. The assumption may be wrong.

Wilma


"bruno de baenst" wrote in
:


"Wilma" schreef in bericht
. 125.201...
I'd put my money on Deep Fritz. I had CM9k analyze two of the games
from his match with Topalov, Games 9 and 10. He did not play with
great precision in either game with the settings I used, 4Mb hash,
180 seconds per move. Compared with what CM says is the best move,
Kramnik scored 53% in Game 9 which he lost and 64% in Game 10 which
he won.

If these two games are representative, he may not be playing with
sufficient precision to beat a GM program.

Wilma

I don't really get the point, because Kramnik doesn't play the same
moves as the pc he plays worse? Maybe he just plays better moves?

bruno de baenst wrote:
"Wilma" schreef in bericht
. 125.201...
I'd put my money on Deep Fritz. I had CM9k analyze two of the games from
his match with Topalov, Games 9 and 10. He did not play with great
precision in either game with the settings I used, 4Mb hash, 180 seconds
per move. Compared with what CM says is the best move, Kramnik scored 53%
in Game 9 which he lost and 64% in Game 10 which he won.

If these two games are representative, he may not be playing with
sufficient precision to beat a GM program.

I don't really get the point, because Kramnik doesn't play the same moves as
the pc he plays worse? Maybe he just plays better moves?

Not necessarily. Sometimes a GM plays much better moves, but sometimes
they can make a human tactical blunder. Computers never make that sort
of tactical oversight error, although they can be driven into subtle
positional traps by a well prepared expert. And if the game reaches an
endgame at 6 or fewer pieces on the board the computer is exact.

I don't know what features CM offers, but the blunder check feature on
Fritz 9 or Shredder 10 has no trouble at all in findng and highlighting
glitches in the play of either GM in the current WCC match. It shows
how close to super-GM strength the current engines are.

I wouldn't trust an engine to choose the absolute "best" mover either.
Allowing about 5-10 centipawns of slack in the evaluation function for
annotation is probably about right.

Even at rather shallow levels of fixed depth (with quiescence) Shredder
10 can find significant errors in the play of games 8,9,10. A GM would
utterly hammer the engine into the ground at these settings but it can
still find key tactical blunders.

Taking the recent game 11 draw as a concrete example with no clear
blunders analysis by Shredder 10 shows that

29. f5 loses the advantage for white built up so steadily up to that
point

29. Kf3 or 29. e3 or 29. Bxe6

All look to be better by nearly half a pawn even at very deep levels of
analysis
(well beyond what a computer might manage in real time match play)

In one of the other games at a pivotal point there were about 8
stronger moves than the one actually played in the match (again using
overnight deep positional analysis).

It isn't safe to assume the computers idea of the best move is truly
the best. But it is usually safe to assume that if the computer sees a
move played that evaluates as more than 3 standard deviations weaker
than the best it is worth doing much deeper analysis.

Regards,
Martin Brown

wrote:
Wilma wrote:
I used, 4Mb hash, 180 seconds per move.

Only 4 MB hash for analysis? That's nowhere NEAR enough. For 3 minutes
per move (assuming a reasonable fast computer, let's say a 2.4 GHz) you
should have 24 MB of hash to be safe.

Here's how to test how much hash you need. Find any reasonably complex
middle-game position and turn on the Mentor window. Count how many
positions are analyzed in 10 seconds. That is how many BYTES (rounded
upward to the nearest MB) you would want to allocate if you were going
to analyze for 10 seconds per move.

That calculation is meaningless.

Firstly, each hash table entry must take much more than one byte. At
the very least, it has to be storing the hash key (presumably 64
bits), the recommended move (say, another 8 bits), the evaluation of
that move (say, 16 bits) and the depth to which that evaluation has
been made (say, 8 bits). So that means that each entry is taking
about twelve bytes, not one.

Secondly, hash tables are essentially random beasts so there's very
little chance that, if you store a million positions in a hash table,
you'll be able to get them all out again. If you're going to be
looking at a million positions, you need a hash table with many more
entries than that to be able to retrieve a large proportion of those:
allow at least a factor of two.

Thirdly, if the hash table isn't being cleared between moves, the
bigger your hash table is, the greater is the chance of being able to
retrieve positions that the engine thought about last move. So again,
bigger is better.

I've yet to see a convincing argument that the correct size to use for
the hash table is anything other than ``as much as you can fit in RAM
without swapping.''


Dave.

--
David Richerby Flammable Gerbil (TM): it's like a
www.chiark.greenend.org.uk/~davidr/ children's pet but it burns really
easily!

Wilma wrote:
I'd put my money on Deep Fritz. I had CM9k analyze two of the games from
his match with Topalov, Games 9 and 10. He did not play with great
precision in either game with the settings I used, 4Mb hash, 180 seconds
per move. Compared with what CM says is the best move, Kramnik scored 53%
in Game 9 which he lost and 64% in Game 10 which he won.


I guess the validity of the computer analysis depends upon whether Kramnik can
beat CM9k or not. If so, then perhaps Kramnik should be analyzing whether
CM9k can beat Topalov. ;-)

--
Thomas T. Veldhouse
Key Fingerprint: 2DB9 813F F510 82C2 E1AE 34D0 D69D 1EDC D5EC AED1