Computer Chess Club Archives




Subject: Re: is the

Author: Bruce Moreland

Date: 15:16:51 07/20/98

Go up one level in this thread

On July 20, 1998 at 17:21:31, Tom Kerrigan wrote:

>I'll do the test, but I'm not sure why.
>The test that I did ran for half an hour, which is enough to average out any
>"warming up" necessary, in my opinion. I also ran the unlinking program first,
>so the "warm up" advantage would be with the dead bit program.
>It also seems to me that the code has been changed enough and the programs
>probably ran in different memory spaces anyway, so I doubt caching played a
>major role in this.

If you do it this way, sandwiching versions A and B, and running the thing
several times like ABABABABAB, I believe that you can make a pretty good
conclusion about what is happening.

You can get a clue about the consistency of the timings you are getting, since
you can compare the A's with other A's, and B's with other B's, and you can also
show that you aren't having initialization effects, and you are less likely to
make a conclusion from a single run that might have gone bogus.

If you show that A is slow for one run, then the immediately following B is
fast, then the following A is slow again, and the following B is fast again, ad
nauseum, I would think that you've pretty much proven that B is faster than A.

Whether or not this demonstrates the practical superiority of one algorithm over
its variant is not as clear, but it's certainly evidence.

One other obvious question is, are you sure that A and B do exactly the same
thing?  Since the two versions should search exactly the same tree, do you have
some way of checking that this is true?  What I do in these cases is make sure
that the tree sizes are exactly the same after each ply in each problem of the
test suite.  This isn't perfect but it seems to be close enough.


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.