Computer Chess Club Archives




Subject: Re: New tool to estimate the statistical significance of match results

Author: Joachim Rang

Date: 13:43:56 07/17/04

Go up one level in this thread

On July 17, 2004 at 16:28:26, Rémi Coulom wrote:

>On July 17, 2004 at 14:21:00, Joachim Rang wrote:
>>Hi Remi,
>>that is exatly what I need right now, thank you very much.
>>Unfortunately the MonteCarlo calculation takes really long if I feed in the
>>result of two 400 games-gauntlets of two different version. I aborted after 10
>>minutes. Any idea how long it might take to compute on my Athlon xp 1540 Mhz? Is
>>ther a way to make it a bit less accurate but quicker?
>>regards Joachim
>I ran a few experiments with 400 games, and it keeps displaying a value of
>-1.#IND00, which indicates a division by zero. The program seems to work OK with
>200 or 300 games, but not with more. Beyond 300, exponents become too high for
>the floating point precision. I may try tomorrow morning to replace the built-in
>"long double" type by a more accurate type of the GNU multiprecision library, so
>that p^400 is not equal to zero for probabilities p  < 1. That should make the
>program applicable to bigger matches.
>What are the numbers of your gauntlet matches ?

I look forward to a new version.

The two gauntlets I want to compare have the folowing numbers:

+197,= 95,-108
+189,= 81,-130

regards Joachim

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.