Author: Christophe Theron
Date: 11:30:28 10/09/01
Go up one level in this thread
On October 09, 2001 at 13:56:45, Dann Corbit wrote:
>On October 09, 2001 at 11:56:27, Christophe Theron wrote:
>>On October 09, 2001 at 05:47:25, Dann Corbit wrote:
>>>On October 09, 2001 at 05:19:00, Uri Blass wrote:
>>>>On October 08, 2001 at 05:48:55, Daniel Clausen wrote:
>>>>>Hi
>>>>>
>>>>>On October 08, 2001 at 03:13:22, Jouni Uski wrote:
>>>>>[snip]
>>>>>
>>>>>>In average 3,70. This gives (log2/log3,7)*84 = 45 points for speed doubling.
>>>>>>Previously it has been spoken about 60-70 points for doubling, so there is
>>>>>>diminishing returns now!
>>>>>>
>>>>>>Jouni
>>>>>
>>>>>That or either the previous or the new claim is based on shaky data. ;)
>>>>>
>>>>>Sargon
>>>>
>>>>In reply to Dann Corbit I can say that the error here is smaller than the error
>>>>of one program because when we look at the average gain of a lot of programs we
>>>>get smaller error than the erro in one program.
>>>>
>>>>I guess that the error margin for the average is close to 10 elo.
>>>
>>>Unless, of course, it is larger.
>>>There are a lot of variables, all moving at once.
>>>You might be right. But if you are right, it is because you are a good guesser.
>>
>>Uri prefers solid data over guesses, but when he has to guess in general he is a
>>good guesser. :)
>>
>>Anyway it is true that even +/-10 elo on this matter is a HUGE error bar. We are
>>trying to detect a change of 10 elo in improvement from doubling the computer
>>speed.
>
>60-10=50
>45+10=55
>
>With a 10 ELO error for a single standard deviation, the answers overlap.
>As others have pointed out, the number of draws found in the games for the
>fastest CPU's (in comparison with the same programs on slower CPU's) might be
>examined for statistical relevance.
Yes.
Christophe
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.