Computer Chess Club Archives




Subject: Re: Fairly good examples of eval inconsistencies between Rybka versions...

Author: Uri Blass

Date: 02:34:30 01/02/06

Go up one level in this thread

On January 02, 2006 at 04:26:55, Vasik Rajlich wrote:

>On January 01, 2006 at 17:06:43, enrico carrisco wrote:
>>I stated here in CCC days after Rybka was released that the nps was manipulated
>>but lacked any real convincing examples of this.  I even ran the idea by Chrilly
>>when he was disassembing Rybka, but he had no comment on it.
>>In the following position, we see two different evaluations for Rybka 1.0 Beta
>>64-bit and Rybka 1.01 Preview 2 64-bit.  Of course, without knowing what
>>undocumented changes (if any) were made from 1.0 Beta to Preview 2, it's hard to
>>deduce anything concrete.  However, Vasik has posted here a few times that
>>nothing was changed with the strength or eval (I believe...  Not trying to quote
>>Specifically, the time-to-ply figures are odd.  Preview 2 appears to slow down
>>in time to ply while Rybka 1.0 Beta seems to 'speed up' as the depth increases
>>(from nps figures.)
>>The second position, below, shows the differences in nps better -- where the
>>qsearch demands a little recognition...  :)  Time-to-ply is very close (unlike
>>the first position) but the nps differs substantially...
>I already explained this here - did you really miss it?

I remember that you said that you changed the definition of nodes.
I think that the reason was simply that your old definition did not count nodes
in the qsearch.

I think that node should be a move that you make inside your data structure(it
is not important if you call it makemove or call it another function inside the
It make sense to have slow evaluation but it does not make sense to search less
than 10 nodes per second and I proved that rybka beta can in extreme conditions
search less than 10 nodes per second.

I see no logical explanation to what I see except assuming that you search
captures in the beta version and do not count them as nodes.

I have no problem with hiding the number of nodes and I think that there is no
problem if you always report 1234567 knodes (people will understand that it is a
joke and you simply do not want to discover the number of nodes).

Alternatively you can divide the number of nodes by 1000000 and people will also
understand that it is a joke.

The problem is that I do not like giving misleading information and the number
of nodes that rybka reported did not seem to be a joke when I first looked at it
and only after trying an extreme position when there are many captures I found
that it is the case.


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.