Computer Chess Club Archives




Subject: Re: Proving something is better

Author: Omid David Tabibi

Date: 21:53:55 12/18/02

Go up one level in this thread

On December 19, 2002 at 00:27:53, Miguel A. Ballicora wrote:

>On December 18, 2002 at 11:07:49, Omid David Tabibi wrote:
>>On December 18, 2002 at 03:21:02, Bruce Moreland wrote:
>>>On December 17, 2002 at 20:44:45, Omid David Tabibi wrote:
>>>>Heinz' experiments showed that std R=3 is weaker than std R=2 [1]. Bruce's
>>>>Ferret also used std R=2 in WCCC 1999 [2]. So I took the one which is believed
>>>>to be stronger (std R=2), and showed that vrfd R=3 is superior to it.
>>>Yes, but it is possible that normal R=3 is stronger than R=2, and that your
>>>enhancement is weaker than R=3.
>>>You directly claim to be better than R=2, which is acceptable, but you imply
>>>that you are better than R=3.  It is possible that you are better than R=3, but
>>>you have not shown this to be true.
>>>You could have anchored your conclusion much better by demonstrating that your
>>>algorithm is superior to R=3 as well.  It's important to do this, since your
>>>algorithm is related to R=3.
>>>Whether my own program uses R=2 or R=3 has nothing to do with this.  That R=2 is
>>>accepted convention is all the more reason to challenging it by investigating
>>>R=3.  If yours is better than R=3, you are winning on all fronts.  If it is not
>>>better than R=3, your algorithm is very suspect, since it behaves differently
>>>than expected.  Even if it's already *proven* that R=2 is better (which I
>>>doubt), you should take the time to prove it here, because if you prove it again
>>>it's evidence that your program is operating properly.
>>>It's nothing personal.  I would argue these points regardless of who wrote the
>>Have you ever conducted any research? If so, you would have known that a
>>researcher doesn't examine everything since the creation of earth, he takes
>>something which is known to be better and tries to improve it.
>In experimental sciences, many times things are repeated to certify that the
>rigth conditions for measures are correct. Many times, those serve as controls.
>It pretty much depends.

True. If you repeat published experiments and your results simply confirm them,
there is no point to publish, but if your results contradict them, then you have
a new case.

Before starting the experiments on verified null-move pruning, I tested R=2
against R=3, and R=2 fared better. A few months ago I posted those results, also
claiming that in longer time controls the superiority of R=2 over R=3 is not
that significant (nevertheless, still superior).

But the main point of the article isn't comparison between R=2 and R=3. It is
about showing that vrfd R=3 is superior to both R=2 and R=3, and the
experimental results conducted on thousands of positions strongly confirm that.

For example, see Tables 2 and 6: vrfd R=3 solves about the same number of
positions as std R=1. See Table 4: vrfd R=3 solves far more positions than R=2
and R=3.

Based on these results, there is no room for doubt as to vrfd R=3's superiority.

>>I didn't think that someone will seriously claim that std R=3 is better than std
>>R=3; but now, I'd be glad to write another paper comparing those two, and also
>>mentioning fixed time comparisons if people find it interesting. Because
>>although not appearing the article, I have conducted tens of other types of
>>experiments (including fixed time) and I _know_ that vrfd R=2 is clearly
>>superior to std R=3.

This page took 0.04 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.