Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Some analysis of Deep Fritz for kasparov-deeper blue first game

Author: Vincent Diepeveen

Date: 09:30:30 05/07/01

Go up one level in this thread


On May 07, 2001 at 09:57:12, Robert Hyatt wrote:

>On May 07, 2001 at 06:15:24, Vincent Diepeveen wrote:
>
>>On May 06, 2001 at 16:11:11, Uri Blass wrote:
>>
>>>On May 06, 2001 at 14:44:40, Paul wrote:
>>>
>>>>On May 06, 2001 at 14:26:52, Uri Blass wrote:
>>>>
>>>>>On May 06, 2001 at 14:25:58, Uri Blass wrote:
>>>>>
>>>>>>On May 06, 2001 at 14:16:04, Paul wrote:
>>>>>>
>>>>>>>On May 06, 2001 at 14:01:22, Terry McCracken wrote:
>>>>>>>
>>>>>>>>On May 06, 2001 at 05:40:02, Uri Blass wrote:
>>>>>>>>
>>>>>>>>>On May 06, 2001 at 03:51:47, Paul wrote:
>>>>>>>>>
>>>>>>>>>>On May 06, 2001 at 02:28:14, Uri Blass wrote:
>>>>>>>>>>
>>>>>>>>>>>I gave Deep Fritz to analyze similiar number of nodes to Deeper blue and Deep
>>>>>>>>>>>Fritz seems to be clearly better in tactics.
>>>>>>>>>>>
>>>>>>>>>>>Deep Fritz needs only 191728 knodes to see the line Rf5+ Ke3
>>>>>>>>>>>It means only 1 second if I asuume 200,000,000 nodes per second.
>>>>>>>>>>>
>>>>>>>>>>>I believe that Rf5+ failed low at depth 17 for Deeper blue for the reason Ke3.
>>>>>>>>>>>The pv of deeper blue at smaller depthes is Rf5+ Ke2
>>>>>>>>>>>
>>>>>>>>>>>Deep Fritz probably does better extensions than Deeper blue because Deep Fritz
>>>>>>>>>>>see big fail low at depth 16.
>>>>>>>>>>>
>>>>>>>>>>>Deep fritz also can see another fail low for Rg8 at depth 22 when deeper blue
>>>>>>>>>>>could get only depth 17 after similiar number of nodes.
>>>>>>>>>>>
>>>>>>>>>>>I do not believe that you lose more than 2-3 plies from null move pruning(my
>>>>>>>>>>>test suggest that you do not lose even 1 ply at small depthes so I guess that
>>>>>>>>>>>Deep Fritz can search deeper because it is a better software.
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>kasparov - Deeper blue
>>>>>>>>>>>4r3/8/2p2PPk/1p1r4/pP2p1R1/P1B5/2P2K2/8 b - - 0 1
>>>>>>>>>>>
>>>>>>>>>>>Analysis by Deep Fritz:
>>>>>>>>>>
>>>>>>>>>><snip>
>>>>>>>>>>
>>>>>>>>>>I remember from the time this match was played that this was due to some bug
>>>>>>>>>>which was subsequently corrected, so there's not much sense in discussing this
>>>>>>>>>>position. Any other will do, but not this one. Even my program finds Rf5+ in
>>>>>>>>>>seconds.
>>>>>>>>>>
>>>>>>>>>>Groetjes,
>>>>>>>>>>Paul
>>>>>>>>>
>>>>>>>>>I know about the bug
>>>>>>>>>
>>>>>>>>>I am not talking about finding Rf5 but about finding the reason that Rf5+ is
>>>>>>>>>losing.
>>>>>>>>>
>>>>>>>>>I mean to find the fact that line Rf5+ Ke3 that is good for white.
>>>>>>>>>Deeper blue could not see it at iteration 16 and the logfile suggests Rf5+ Ke2
>>>>>>>>>
>>>>>>>>>Deeper blue had a fail low at iteration 17 and the logfile does not give a line
>>>>>>>>>for Rf5.
>>>>>>>>>I guess that it failed low because of Ke3.
>>>>>>>>>
>>>>>>>>>The point is that Deeper blue is slower than top programs in failing low.
>>>>>>>>>
>>>>>>>>>Deeper blue could not find Rf5+ Ke3 after 73 seconds when Deep Fritz can find it
>>>>>>>>>in a few minutes on p800 and it means that it could find it in less than second
>>>>>>>>>if it could search 200M nps.
>>>>>>>>>
>>>>>>>>>Uri
>>>>>>>>
>>>>>>>>I think you're right Uri if you dragged out Deep Blue of May 1997 or at least
>>>>>>>>close.
>>>>>>>>But you have to remember it was 1997 and if my memory serves me correctly,
>>>>>>>>Kasparov was very puzzled by Rf5+ as programms just didn't look at this move
>>>>>>>>at that time, except for Deep Blue in such a short time frame.
>>>>>>>>At least that's what I remeber from what Kasparov mentioned in his notes. Not
>>>>>>>>the exact words, I'm not quoting what Kasparov said or wrote but just what I
>>>>>>>>remember from that time, on the "Old Club Kasparov" hosted by IBM which is long
>>>>>>>>gone.
>>>>>>>>I also remember after many hours of analysis with "computers of the day"
>>>>>>>>P6-200's, that Kasparov finally "understood" why Deep Blue played the "Human"
>>>>>>>>looking move, Rf5+.
>>>>>>>>Actually, I think Kasparov wasted too much energy trying to understand Deeper
>>>>>>>>Blue which I believe exhausted him. Hence, his less than stellar preformance
>>>>>>>>in Game 2 and the rest of the match.
>>>>>>>>
>>>>>>>>Terry McCracken
>>>>>>>
>>>>>>>Deep Blue didn't play Rf5, it played Rd1! Uri wants to analyze the evaluation
>>>>>>>out of the log of Deep Blue ignoring the bug. Seems impossible to me. :)
>>>>>>>
>>>>>>>Paul
>>>>>>
>>>>>>I believe that the bug happened only after failing low on Rf5.
>>>>>>
>>>>>>The logfiles do not give a score for Rd1 so I believe that the bug is not
>>>>>>relevant for Rd1
>>>>>
>>>>>I mean of course is not relevant for the lines before Rd1
>>>>>>
>>>>>>The last line of the logfile of game 1 begins with
>>>>>>11(6)[Rf5](-260)v [find a move]
>>>>>>
>>>>>>I never see the words find a move in deeper blue in other cases and it suggests
>>>>>>that the bug happened only after Deeper blue failed low.
>>>>>>
>>>>>>The line for Rd1 is also a short line when previous lines are long lines so I
>>>>>>trust the analysis of deeper blue at depth<17.
>>>>>>
>>>>>>Uri
>>>>
>>>>Uri, you don't know what the bug was, I don't know what the bug was, and now out
>>>>of all the positions you could analyze, you pick the one where it's certain that
>>>>there is a bug in it. Why?
>>>>
>>>>Seems to me there are hundreds of other positions more useful & interesting for
>>>>analysis, for example from the (by Deep Blue at least) excellently played game
>>>>6? But if you want to ... go on by all means. :)
>>>>
>>>>Greetings,
>>>>Paul
>>>
>>>I already analyzed another position from the first game and unfortunately nobody
>>>responded to the analysis and the discussion was about the question if Deeper
>>>blue did processing at the root.
>>>
>>>see http://www.icdchess.com/forums/1/message.shtml?167877
>>>
>>>Uri
>>
>>Uri, for all programmers you had proven the point very clearly.
>
>
>No, he didn't "prove" anything at all.  You can't take output from a program,
>output that is incomplete due to the way they produce their PV, and then try to
>draw conclusions from it.  Any more than you can take evaluations and draw
>conclusions from them without any idea of what the evaluation looks like
>internally...

It's hard to argue with scores produced Bob!

Also it fits into the picture. I remember once you posted that they
tuned their search parameters using 1 ply searches using automatic
testing & tuning.

That never worked for me actually. Not a single machine can tune my
parameters like i can tune them! And i tune myself pretty when compared
to what is possible!

However that sort of tuning makes more sense to me
when using preprocessor tables!





This page took 0 seconds to execute

Last modified: Thu, 15 Apr 21 08:11:13 -0700

Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.