Author: Peter Fendrich
Date: 09:42:52 07/30/98
Go up one level in this thread
On July 30, 1998 at 00:48:21, Shaun Graham wrote: > How are computer chess program ratings calculated, what is the formula? >It has occurred to me that perhaps it makes no sense to calculate computer >ratings in the way that human ratings are calculated. It's hard to describe my >point so i will give an example. You're absolutley right about that! The reason is that a chess program on a specific hardware has the same playing strength during all time. That's not the situation for humans. Our strength is varying during our life time. The original ELO formula is suited for the human situation. The rating list computed by SSDF is taking care of this and doing what you're proposing, namely computing the rating from scratch for all programs every time. The "only" problem to deal with is that none of the programs have any fixed rating to start with but that is handled by an iterative algoithm. It's a little bit more complicated than your example below but about the same idea. //Peter > >A human rated 2200(base rating) plays a tournament of ten games, after the >tournament his rating is calculated to be 2280. Arriving at this calculation >his initial rating of 2200 was used. In his next tournament of ten games his >rating is calculated to be 2320. To arrive at this rating 2280 was used in the >base calculation. This seems to work for humans because you can hypothesize >improvement thus a new base rating works, but for computers it seems as if >having a new base rating may be problematic because their is no so called >"improvement". In a way it seems to me that it all works fine and I know the >rating is supposed to just measure performance but i just have an intuitive feel >that there should be a difference in the calculations. > >Another problem that i have is that say i have 50 tourney games played, and i >calculate the rating in ten game increments. I use the new rating achieved >after each ten as the base to do my calculation. my final rating is different >than if i acted as if all the games happened in one tournament. Perhaps i'm >screwing the calculation? > > >I have hypothesized that to calculate the rating of a program based upon all the >games played at once you should use the provisional formula > >For example Chessmaster 5555 on Kai Luebke's web page >"http://www.geocities.com/TimesSquare/Ring/8448/chessindex2.html" in 40/2 games >has won 29 games, drawn 13, lost 11 games. I gave a flat rating of 2400 to all >opponents > >the calculation went as follows [(29wins * 2400)+(29wins*400)] + [(13draws * 0)+ >(13draws * 2400)] + [(11Loss * 2400)-(11Loss * 400)] = 2535(Chessmaster is a >MONSTER) > >This is the standard uscf formula of adding 400 points to opponents ratings for >wins, and subtracting 400 for losses, then dividing by the number of games >played for a provisional rating. > >The opponents were Rebel9 & 8, Mchess 7, Fritz 5, Hiarcs 4 & 6, crafty 14.8, >Junior 4.6, and Genius 5.
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.