Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: New SSDF-list

Author: Robert Hyatt

Date: 06:37:35 02/24/98

Go up one level in this thread


On February 24, 1998 at 09:34:46, Robert Hyatt wrote:

>On February 24, 1998 at 06:24:04, Thorsten Czub wrote:
>
>>I thought fritz/nimzo vs. Rebel are an exception ! :-)
>>From my point of view there is NO pattern in the games - rebel8/9 looked
>>pretty outsearched ! And this means you don't get a typical pattern like
>>:
>>your king safety is too weak or enemy handles the pawns better or
>>whatever.
>>
>>Concerning the hash-tables I have to admit that fritz seems to handle
>>them different than other programs. fritz is very hash-table sensitive.
>>I don't think other programs behave the same way concerning hash like
>>fritz does.
>>Maybe Frans has programmed something different.
>>You can test yourself by reducing hash-tables in fritz5 and find out
>>THAT fritz is very sensible if you reduce the hash.
>>
>>no - the fast searchers have - for the first time in the history of
>>computerchess and ssdf-list, overtaken the knowledged based programs.
>>THATS the big surprise.
>
>
>you haven't been around long enough to say that.  Can you spell Chess
>4.x?
>they went from smart/selective to dumb/fast, and set the world on fire.
>Cray Blitz went from smart/selective to dumb/fast (at first) and went
>way
>up in speed, with a real rating of 2258 (USCF) in 1980.  There's nothing
>new here at all.  You simply find that there are two approaches.  Fast
>means simple, which translates into easy debugging issues.  Smart means
>complex, which translates into complex debugging issues.  Although I
>don't
>agree that Rebel is a slow/smart program when it is faster than Crafty
>on
>equal hardware...
>
>
>
>>When we played in Paderborn we had many discussions concerning
>>OUTSEARCHING the opponent, due to the fact that ChessTiger and Nimzo98
>>behaved like this.
>>Nimzo98 outputs did not show this but the games looked like this. Any
>>opponent of ChessTiger was starring with wide eyes on it's search
>>depths. Vincent ran arround and told anybody stories about SCHACH3.0. In
>>the game vs. clever+smart e.g. this was very easy to see. Clever got a
>>fail-high, and after 2 minutes a bad fail-low following Tigers moves.
>>Nimzo98 attacked Tiger like hell, the game turned arround several times.
>>But - in the end was also a draw.
>>
>>Maybe Fritz5 with big hash can OUTSEARCH the others, meanwhile fritz5
>>with LESS hash cannot do this.
>>I have tried it out myself by putting enough ram into my machine. The
>>changes were drastic. WHY ?
>
>
>If I were Frans I'd be concerned that my hash-replacement algorithm was
>*not* doing its job.  IE in normal positions, we only get 10-20% hash
>hits
>anyway.  So there's no reason to go in the dumper when the table fills
>up,
>*unless* the hash replacement algorithm is not working well, or is too
>simple
>to work well..
>
>
>>WHat is Frans doing with the hash-tables ?Or better - what is he doing
>>different that increasing of hash lets the playing-strength increase
>>linear !
>>One main thing of course is the big number of HITS fritz gets with
>>computing almost 200K NPS.
>
>but it isn't that big.  IE Cray Blitz, searching 5-10M nodes per second
>reports 20-30% hash hits (max) in the opening/middlegame.  It does go
>*way*
>up in endings of course...

there is another idea I just thought of.  Suppose Fritz is very sloppy
in
move ordering (IE it doesn't try the history heuristic, to avoid the
extra
instructions needed).  In that case, it might be *very* dependent on
getting
a transposition/refutation table hit to suggest the best move to try.
This
would be *very* ugly, but it is possible...



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.