Author: Sune Fischer
Date: 08:43:26 02/27/02
Go up one level in this thread
On February 27, 2002 at 11:07:36, Uri Blass wrote: >>>The rating is dependent in the opponet that the perfect player chooses to play. >> >>No it is not, look at the formula, it is a normal distribution. >> >>>The perfect player may get 100% against my program on p800 because my program is >>>a deteministic program that always does the same mistake so if you assume the >>>perfect player plays only against my program then the perfect player is going to >>>get infinite rating. >> >>Your program is deterministic by your own words, so must score even worse than >>one doing random moves. > >it is going to score worse than random moves against the perfect player but it >scores clearly better against a lot of other players. Actually if it is deterministic, it will not score very high on a chess server, someone will beat it once and repeat the game over and over and nick all its rating. This is why you need a book or something to randomize it just a little bit. >>>The perfect player may get 100% against a player with a rating of 2000 when the >>>same player is going to fail to get 100% against a player that is clearly weaker >>>but not deterministic. >> >> >>Please do not ignore the small differences in probability, they are important. >>A 2000 elo player may be beaten by 10^30:1 and a 1000 elo player by 10^35:1, it >>should all add up to the same rating for the perfect player, that is how the elo >>table works. > >The problem is that rating is something that is dependent on the opponents that >you play. > >The perfect player may get 100% against one player with rating 2000 and only >99.9999999999% against another player who has today lower rating based on the >rating system. > The perfect player will _not_ get 100%, that is point. If the perfect player scores 100% he will have an infinite rating nomatter what the elo is of his opponent. If he's playing a deterministic engine, this will actually happen, I don't know if this means the program has minus infinity or the opponent infinity, whatever it is a strange example. So, you must assume less than 100%, even against the worst possible opponent. >You cannot calculate rating for a player without knowing the opponents and their >rating. Actually you can in a way, you just find the elo-difference, it is expresed by: DeltaELO=-400*log(1/p-1) where p is the probability of a win. You just add the opponents rating to this elo. I'm not sure if that is the exact formula used for chess, but I think it is something similar. >My program may get worse rating than the random move generator if they play >enough games against the perfect player but if you give both of them to play >against beginners then it is going to get better rating. If it is deterministic, it will in principle lose all matches after it has lost just one, this will not happen to a random program. >Uri
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.