Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Verified Null-Move Pruning, ICGA 25(3)

Author: Omid David Tabibi

Date: 12:55:56 11/26/02

Go up one level in this thread


On November 26, 2002 at 15:23:54, Peter McKenzie wrote:

>On November 26, 2002 at 12:26:58, Omid David Tabibi wrote:
>
>>On November 26, 2002 at 12:18:10, Omid David Tabibi wrote:
>>
>>>On November 26, 2002 at 07:09:07, Gian-Carlo Pascutto wrote:
>>>
>>>>On November 25, 2002 at 22:32:28, Dave Gomboc wrote:
>>>>
>>>>>>No. Maybe I shouldn't have mentioned it - it's not really fair
>>>>>>to Omir as he does publish his stuff.
>>>>>
>>>>>In that case, I'll take it as a courtesy if you don't criticize any work I
>>>>>happen to publish because it uses node counts instead of wall clock timings.
>>>>
>>>>I don't think it's fair to say 'what you published is crap, I have
>>>>something much better but I'm not telling you what and I'm not going
>>>>to publish it or post test results from it'.
>>>>
>>>>I think it's fair to say 'what you published is crap because you did
>>>>not test it correctly and you compared only to inferior methods whereas
>>>>better methods were already known and published'
>>>>
>>>>I don't think it's fair to criticise Omir because his scheme does
>>>>not work in my engine. I think it's fair to criticise him because
>>>>he did not include Heinz scheme in his tests.
>>>>
>>>
>>>What do you mean by "he did not include Heinz scheme in his tests"?
>>>
>>
>>Oh, if you mean adaptive null-move pruning, I didn't consider it since its
>>tactical strength is not more than standard R=2.
>
>I'm not exactly sure what you mean by 'tactical strength'.  I suspect you mean
>'can solve more test suite positions at the same depth'.  That can be useful for
>certain comparisons but isn't a good absolute measure of how good an algorithm
>is.
>
>Assuming that verified nullmove is tactically stronger than standard R=2, then
>it must be tactically stronger than adaptive nullmove (at the same depth).  Does
>this make it better?  Of course not because adaptive nullmove may, for example,
>use significantly less nodes at large depths.  If this is the case (I don't know
>if it is) then another comparison method may need to be used.
>
>The main thing is that it is accepted that adaptive null-move pruning is better
>in real games than standard R=2, and surely that is the most important thing?
>Test suites are a means to an end, not an end in itself!
>
>Personally, I am only interested if verified nullmove is better than adaptive
>nullmove in real games.  Now it seems you have shown that for your program
>verified nullmove is likely better than standard R=2 in real games (because it
>is tactically stronger, AND has smaller trees), which is interesting but not
>really crucial because we already have something else better than standard R=2.
>
>Well done on writing the paper, it has certainly stimulated interest which is a
>very good thing.  But a comparison between adaptive nullmove and verified
>nullmove would have made it much more interesting.
>
>Regards,
>Peter
>
><snip>

In my tests verified null-move pruning constructed a smaller tree than adaptive
null move pruning.

It is reasonable since the "backbone" of verified null-move pruning is a fixed
reduction of R=3, while the adaptive one uses both R=2 and R=3. So it is
reasonable that on every program starting from a certain depth adaptive
null-move pruning will always construct a smaller search tree.






This page took 0.1 seconds to execute

Last modified: Thu, 07 Jul 11 08:48:38 -0700

Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.