Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Rybka improvements

Author: Vasik Rajlich

Date: 04:01:15 01/27/06

Go up one level in this thread


On January 26, 2006 at 05:07:17, Keith Hyams wrote:

>On January 26, 2006 at 01:59:48, Kurt Utzinger wrote:
>
>>Has anybody ever seen better overall scores with the many
>>Rybka beta's compared with the first Rybka 1.0 Beta 32-bit.
>>[http://www.husvankempen.de/nunn/cegtrating4040best.html#4]
>>[http://www.utzingerk.com/rybka10beta_test.htm]
>>I can't remember and am therefore asking if we can indeed
>>speak of any progress (apart from reducing number of bugs).
>>It would be a pleasure to see a list with some hundred games
>>vs other engines (no dummy bullet 1m+1s and 3m+0s games)
>>proving the contrary of my assumption.
>>Best regards
>>Kurt
>
>If I play a bullet game between Rybka and another engine I obtain information
>about which of the 2 engines has the more effective  combined search and
>evaluation to a search depth of 9 ply. If you play a time limit of 1 hour you
>get the same information down to 12 ply.
>
> In the time that you have taken to finish your single game, I have played 40
>games testing Rybka against 40 different engines with 40 different openings.
>
> Please explain why this is a dummy thing to do. Please don't tell me that it is
>because people can't play decent bullet chess because I know this. I also know
>that my computer is not a human and so this is not a valid thing to say.
>
>As you probably know, most of the releases of beta versions of Rybka were not
>designed to increase strength. My opinion on the strength increase between the
>original and beta 12 is that beta 12 is between 30 and 40 Elo stronger. This
>opinion is not worth much because it is based on dummy games.
>
>If I am correct about the increase in strength, then this change is quite
>impressive. It has been achieved in about 8 weeks. Other authors of top-end
>engines often achieve less than this in one year.
>                                    Regards
>                                          Keith

Testing at really fast time controls is an interesting idea. It's not really
appropriate (IMHO) for serious rating lists comparing engines, but it's not a
bad way to get a lot of data about engine settings or smaller changes.

The one thing to watch out for is that even though most changes work equally
well at all time controls, if you only test using very fast time controls, you
are guaranteed to gradually skew your program in favor of fast play. There
should always be some sort of balance.

Vas




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.