Author: Dan Homan
Date: 09:56:44 01/18/00
Go up one level in this thread
On January 18, 2000 at 11:45:37, Will Singleton wrote: >On January 18, 2000 at 08:49:48, Vincent Diepeveen wrote: > >>"R(mhz) rating/log(mhz)/100. Shows the performance of an account, >> taking into consideration the cpu speed." >> >>First of all Will is taking into account for his formula a logaritme, >>combines that with rating and computer Mhz. >> >>This assumes a logarithmic connection between Mhz of a machine and its >>rating. >> >>Obviously this is wrong. >> >>Let's take crafty: >> >>Wild 2148 [6] 195 50 8 253 2148 (06-Aug-1999) >>Bullet 3033 6241 1412 1018 8671 3191 (27-Oct-1999) >>Blitz 2858 48312 12919 8716 69947 3268 (27-Oct-1999) >>Standard 2550 3761 1496 1117 6374 2785 (08-Aug-1999) >> >>Machine = 4x400 = 1600Mhz >>rating = 2550 >> >>So this would make >> R(mhz) = 25.50 / log(1600) = 7.95 >> >>Now that sounds a bit low to me personal already. >> >>Let's now assume that Bob wants to top the green list: >> >> He needs 11.0 then as R(Mhz) to top it more or less. >> >>So 11.0 = rating / 100*log(1600) >> rating = 11 * 100 * log(1600) = 3525 >> >>So crafty needs an incredible rating of 3525 to top the greenlist. >>Good luck Bob! >> >>Why invent a formula if it's in advance already dead wrong? >> >>Rating is not based upon the 10th log of something. >>It SURE does not represent performance taking into account >>the machine speed. >> >> >>Vincent > > >Couple things. First, crafty's efficiency isn't 100%, I don't believe. I use a >figure like 75-80% when calculating the effect of multiple cpus, which is just >an approximation. But you're right, the formula seems to work better in a >narrow range. Have a better formula for me? > >But come on, Vincent, the whole list is just an approximation. Many folks feel >that ICC ratings have little real-world significance, due to the factors present >in server play. Again, the list is just for entertainment purposes, it doesn't >purport to be anything official. > >Having said that, on ICC one does see a general correlation between ratings and >perceived strength. > >Will Hi Will, Your adjustment does something to correct for the faster is better rule - and the logrithm accounts for the fact that faster isn't linearly better, but Vincent has a point in that (beyond the above considerations) your adjustment is basically pulled out of thin air. A better approximation might be to use the general rule that doubling the speed gives about seventy elo points. Adjusting for that rule is easy... Adjusted Rating = Rating - 70 x log[2](MHz) - 2000 (where log[2] is log base 2) So a 400 MHz machine would have 70 more poinst subtracted than a 200 MHz machine. The "- 2000" at the end of the formula is arbitrary and I just put it in so that no one will confuse the adjusted elo ratings with the FIDE or USCF scale. An added advantage to this formula is that the end result is still an elo type number, so win/loss probabilities can be judged in the usual way by comparing differences between the adjusted ratings. One could refine this and adjust for the difference between G3/G4 and PII/PIII type processors (maybe replacing the MHz number with a Specint95 for the same machine), but at this level of approximation I am not sure it is necessary. - Dan
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.