[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Bug-gnubg] Results from the bot shootout

From: Achim Mueller
Subject: Re: [Bug-gnubg] Results from the bot shootout
Date: Thu, 5 Jul 2007 14:14:19 +0200
User-agent: Mutt/1.5.9i

* Joseph Heled <address@hidden> [070705 12:41]:
> Even if gnubg wins a match only 49.5%, in a set of 1000 matches there
> is more than 5% chance that gnubg wins 519 of them. that what 95% (one
> sided) confidence interval means.

I guess I got it now. I probably was mislead by an article of Chuck
Bower in GOL were he described std.err as sqrt(a*b/1000) and sigma as
(a-b)/std.err (which is probably wrong).

I just had a chat with Douglas Zare. He says that sigma should be
0.5(a-b)/std.error. In other words, don't take the difference between
both bots, but between a bot and the mean (which is 500 here).

If that's correct - I'm not sure and a bit puzzled now - we get the
following numbers:

gnu/jelly       5.88        100.0%
gnu/snowie      0.69         51.0%
gnu/bgb         1.20         77.0%

bgb/jelly       2.15         96.8%
bgb/snowie      0.25         19.7%

snowie/jelly    4.56        100.0%

This definitely makes more sense. Sorry for the confusion.



reply via email to

[Prev in Thread] Current Thread [Next in Thread]