Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: one general request to CCRL people

Author: Uri Blass

Date: 17:03:49 02/12/06

Go up one level in this thread


On February 12, 2006 at 19:04:55, Heinz van Kempen wrote:

>On February 12, 2006 at 17:20:31, Ray Banks wrote:
>
>>On February 12, 2006 at 15:23:50, Alessandro Scotti wrote:
>>
>>>Hi,
>>>is there a table somewhere that shows how much Elo can an engine get from being
>>>10%, 20%, ... 100+% faster?
>>>Usually speed improvements are considered marginal to "algorithmic"
>>>improvements, but on CEGT the difference between Deep Shredder 9.12 x64 2CPU and
>>>Shredder 9 is 65 elo, and between Rybka 64-bit and Rybka 32-bit an amazing 62
>>>elo... in the latter case not bad at all for just recompiling the program!
>>>I'd rather have my program slow and correct than fast and buggy, but I'm still
>>>wondering if it's still worth to invest some effort in optimizing things a
>>>little.
>>
>>http://kd.lab.nig.ac.jp/chess/CCRL-4040/rating-table-all.shtml
>>
>>We currently have a difference of 42 ELO Rybka 64 bit vs 32 bit. Our time
>>control is more than twice as long as CEGT though, and that could have an
>>impact. Plus our volume of games is fairly low so the statistical error is
>>somewhat higher unfortunately  :-(
>>
>>We are also testing Naum in both 32 bit and 64 bit versions, so that will also
>>give more information. But that will take 2-3 weeks probably
>
>Hi,
>
>the explanation of differences is more likely to be explained from too few
>games, than from giving more time. Even compared with Blitz we have very few
>differences for engines. Going to longer time controls mainly has the effect to
>be unable to test a lot of engines and versions, especially more amateurs.
>
>So CEGT will resist to do that except from tournament time control games not
>used for statistics but for investigation of opening systems and for giving
>some high quality games.
>
>As we have totally different concepts and because of the history all this has
>had I would suggest to avoid any comparison between our groups, at least from
>CCRL people. We did not do this from our side and it would be fair that you also
>avoid it.
>
>Thanks for listening.
>Heinz

I think that the difference may be dependent on the opening position that you
start.
You can do a round robin tournament for every position.

If you have 151 programs you can get 300 games for every one of them in a single
position by round robin tournament when every 2 engines play 2 games.

It will not give you correct rating lists of the engines because engines may be
lucky in specific position that they like but I think that it may be interesting
to find positions when speed is relatively more important and positions when
knowledge is relatively more important.

Uri



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.