# Computer Chess Club Archives

## Messages

### Subject: Re: Proving something is better

Author: Omid David Tabibi

Date: 17:05:48 12/17/02

Go up one level in this thread

```On December 17, 2002 at 17:30:36, Bruce Moreland wrote:

>Omid wrote an article that's in the ICGA that I received today, and there are
>two tables:
>
>D        R=1               R=2            R=3           VR=3
>9     1,652,668,804    603,549,661   267,208,422    449,744,588
>10   11,040,766,367  1,892,829,685   862,153,828  1,449,589,289
>
>D        R=1               R=2            R=3           VR=3
>9        64                62             53            60
>10       71                66             65            71
>
>The first table is nodes taken to search to depth D with four techniques, which
>are standard null-move with various R values, and a "verified" null move search
>with R=3.
>
>The second table is the number of problem solutions in a test suite, given
>particular depths of search and using these techniques.
>
>The conclusion is that VR=3 is better.
>
>Does anyone else see the two problems here?
>
>1) The amount of time taken to process a leaf node may very well be less than
>the amount of time taken to process an interior node (not including its
>children, of course).  So if the tree shape changes, it is possible that it
>could take longer to search a smaller tree.  Unless that is ruled out, all that
>has been proven is that VR=3 works in fewer nodes.  That is pretty interesting,
>but I think it makes a stronger case if *time* is used as well, so we will know
>that there is at least one real case where this technique makes the program
>faster.
>
>2) The amount of nodes traversed to get through ply 10, with R=3, is about 60%
>of the number taken to get through ply 10 with VR=3.  It can be assumed (perhaps
>wrongly, due to my previous point) that this search takes 60% as long.  The
>number of solutions found by R=3 is fewer than with VR=3, granted.  But what is
>the R=3 version doing while the other version is trying to finish up ply 10?  It
>is going on to ply 11.  How many solutions has it found in ply 11 before VR=3
>has finished ply 10?  In this case, 65 is sufficiently less than 71 that the
>answer is probably less then 71.  But maybe not!
>
>It seems likely that VR=3 is tactically faster than R=3, but I cannot know for
>sure, since the results have not been reported!  We do not know if VR=3 is
>tactically *faster* than R=3.  Isn't that an important point, since all of us
>obtain benefit?
>
>I have seen people report results like this forever.  I wish that they would use
>a the more sensible method of reporting number of solutions found correct in a
>certain amount of *time*, since that is the true measure of tactical speed.
>
>Nothing personal, Omid.
>
>I'm trying to verify your results by using ECM with Gerbil.  First I have to get
>good numbers for R=2 and R=3.
>
>By the way, if anyone wants to take Gerbil and use it as a second example when
>writing articles like this, I'm all for it.  It's simple so it should be pretty
>easy to modify.
>

you've raised, several weeks ago (with interesting comments by Robert Hyatt,
Gian-Carlo Pascutto, Tony Werten, Uri Blass, etc). I suggest that you first take
a look at those discussions (check the archives of Nov. 20--30).

Using fixed time instead of fixed depth incurs many problems, e.g., the
experiment will not be repeatable, and will be heavily hardware dependant, in
addition to dependance on engine's NPS. Because of all these reasons fixed depth
experiments are used more frequently for algorithmic comparisons (e.g., see
Heinz' articles as the most recent examples).

Best regards,

Omid.

>bruce

```