Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Fritz5 cooking at SSDF and Nunn test set

Author: Vincent Diepeveen

Date: 18:10:49 02/25/99

Go up one level in this thread


On February 25, 1999 at 11:58:28, Ed Schröder wrote:

>>Posted by Vincent Diepeveen on February 25, 1999 at 07:01:54:
>
>>>It's more likely Rebel9 did not learn because the Swedish test without the
>>>"A" option. Some 6-7 months ago I went through the Rebel9 SSDF games
>>>to verify this. I stopped counting after 100 games. Rebel9's learning (in
>>>auto232 mode) was disabled in all these games.
>>
>>Ed, in that case i would shoot all SSDF members that did it.
>>
>>You provide the option to test without it, and they use it vs fritz.
>>However, the email i received from Karlsson confirms that Fritz5
>>would have disabled learning anyway.
>>
>>>>Even more important than disallowing learning is the fact that only fritz5
>>>>auto232 player can collect the results.
>>>>
>>>>Let me give an example what happened.
>>>>
>>>>I got about 6 months ago a call from a tester who played around 20 games at
>>>>auto232 player against Rebel9.
>>>>
>>>>He told me that i should stop chessprogram development, as my program
>>>>didn't win a single game versus rebel9.
>>>>
>>>>So i first asked whether all games were more or less the same (learning
>>>>from rebel9 you never know...). He told me clearly that the games were
>>>>not the same as he turned off learning.
>>>>
>>>>This confused me. How can one not win a SINGLE Game vs rebel9?
>>>>It was quickly solved when i got back the games.
>>>>
>>>>Diep won lucky several games, but those games were long. The games it
>>>>lost were short (short after book). Diep was mated within say 50 moves
>>>>or something. The games diep won were all 60 moves, after which the
>>>>auto232 game stopped the games. However none of the games diep had
>>>>mated rebel9. Further a big bunch of the games were clear draws, or
>>>>3 fold repetition.
>>>>
>>>>Now diep doesn't collect game results. Rebel does, and the WAY in
>>>>which it does caused the confusion that diep didn't win a single game:
>>>>
>>>>rebel9 screen showed a questionmark where diep won the game,
>>>>and some draws got a questionmark too. Where rebel9 was winning, or
>>>>had mated diep, it showed that rebel won.
>>>
>>>Well Vincent... there are enough people who can confirm the below
>>>description of Rebel9 being in auto232 mode. I think you either have
>>>your facts wrong or you are not so good informed. The text is taken
>>>from the Rebel Home Page.
>>
>>No i am very well informed and what you say confirms this.
>>
>>Games get stopped at 60 moves. Then the games diep is winning are
>>less than -5 for rebel, so rebel gives questionmark.
>
>If games stop at move 60 that's because auto232 has been started
>with /m60. I know only one person who tests like this :-)
>
>And of course if a game is aborted and doesn't have a -5.00 or +5.00
>score the game is marked as "?" (result unknown) and the screen
>isn't updated.
>
>
>>>I think you just forgot to create a "new Rebel database" before you
>>>started auto232. As a result you were investigating old games from
>>>previous auto232 matches too.
>>
>>No ed this didn't happen. What happened was that rebel when winning
>>wins quickly, and when losing needs a long time to lose. That says something
>>about book and toughness of rebel, and the internal evaluation of it.
>>
>>So this mistake was made by reading the screen of your database, and
>>your algorithm exactly reflects why it happened.
>
>Total score from the database always reflects the match score on the
>screen. The score on the screen is nice for an indication but after an
>auto232 session you have to go to all games manually anyway and
>edit all the "?" (result unknown) cases by hand.
>
>I always use the default auto232 setting "/m100". In your case (using /m60)
>the error rate is of course very high as there will be many games terminated
>with scores like +1.xx or -2.xx. And then what? It's not always clear to judge
>such games. Better use "/m100" IMO as it pays off in the end.

Easy to judge for me. don't forget i have over 1000 rating points more
than you. So /m60 is well enough.

>Ed



This page took 0.01 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.