Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Learning programs -Old ones that have learned compete with new programs?

Author: Ed Schröder

Date: 21:54:30 07/08/99

Go up one level in this thread


>Posted by Bertil Eklund on July 07, 1999 at 18:41:26:
>
>>>>Hello!
>>>
>>>What kind of learner is that? Rebel saves the games including the result
>i.e.
>>>0-1. Other programs is aware of the result if the game is SAVED including
>the
>>>game-score. In 99% of the cases the game is saved including the correct
>>>result.
>>
>>Rebel's learner has nothing to do with saving games.
>>
>>At the time of Rebel7 SSDF didn't want to test Rebel7 with auto232 because
>>of the (nice?) build-in automatic abort game function in case of double
>games.
>>Later on request of the SSDF I added the "A" option for Rebel7 which allows
>>double games. The "A" option also manages the auto232 book-learner and
>>position-learner.
>The idea behind this was to greatly expand the opening-books, there was
>programs
>(not Rebel) that only played one or two well-prepared lines and it was
>extremely
> boring for the users. Already Mach3 used a simple but working learner that
>made
>it harder to win the same game over and over again.
>
>>One year ago I took the time to download the Rebel9 games. To my surprise
>>I noticed that most Rebel9 games weren't run with the "A" option. I stopped
>>counting after 100 games. So this is the explanation on your above criticism
>>on Rebel's learner. The truth is you haven't followed your own policy and did
>>test Rebel9 without its learner.

>I play my games with Rebel9 with -a. In the last match against Junior5, 22
>games Rebel9 lost 3 times in 1.c4-c6 2.Nf3 the same variation, no other doubles.

The correct parameter is "a" not "-a".

So REBEL.EXE A

>Or goal is to let every program play at the best possible settings, sorry for
>some sort of misunderstanding. Yes it is also true that one tester by mistake
>used the wrong opening book for some games, but most programs don´t explains
>this very good. In example I have found that Rebel8/9 plays weaker with 60mb
>hash instead of 28mb. Noone has told us the best settings and of course most
>testers tries with 60mb.

The latter indeed could be the case because of a too small hash-key. This
was fixed in the 10c edition.

And thank you for your openness, it is the first time SSDF admits an error.

>Mr de Koning recommended us for his program 16mb is
>more then enough. I have played some games with Rebel10 manually how can I be
>sure that the book-learner works properly, do I have to play until mate?

Playing manually the human book-learner part is active. This is a different
book-learner than the comp-comp book learner which is tuned for long
auto232 matches. If you think a game is over just stop it, it's all taken care
off.

>No this games are for internal use only, several people have asked for
>games and results but they have to relay on the results from Paderborn
>and against humans.

>No harm intended Bertil SSDF

I know.

Ed Schroder




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.