Author: Uri Blass
Date: 08:53:44 01/17/03
Go up one level in this thread
On January 17, 2003 at 11:25:21, Sune Fischer wrote: >On January 17, 2003 at 10:29:26, Uri Blass wrote: > > >>This discussion is not about rating in ICC. >>The fact that the rating on ICC is more correct for new players relative to >>another stupid systems is not relevant for this discussion. > >It isn't, I thought you were complaining about losing rating when winning a >game? > >>>>The rating system was also changed such that people can calculate their rating >>>>because people in general prefer wrong rating that they can calculate and not >>>>right rating. >>> >>>Sure, we must have a rating we can calculate. >> >>Yes but we do not need to have a rating when the players can do it by a simple >>formula. >> >>I think that it is better to have a rating that only a computer program can >>calculate. > >Formula please? There should be investigation to find the best formula but even without investigation I can give you a better formula. For example it is possible to decide that if the last opponent that you play win rating in the next game you get a small bonus of 1/1000 of his(her) gain and if he(she) lose a game you also lose 1/1000 of the loss. The point is that rating is only an estimate and if you see that a player win rating then the estimate was probably wrong so your rating can change. 1/1000 is probably too small and it is possible to find a better coefficient but I suggested a small coefficient only to show that it is possible to get better rating than the rating that is used today. It is also possible to use more than one game of your opponent. players can improve so the weight of later games should be reduced based on the time that the later games are played. > >>>Is there a right rating? I don't think so, certainly not for people who >>>sometimes have a bad day or a good day. >> >>It is possible to get closer to the right rating by a better estimate but nobody >>is interested it. >> >>one example. >> >>It is possible to increase the coefficient(k) of a player that is not a new >>player but does significantly better than the expected result. >> >>suppose that a chess program with rating 1600 play against 1800 players and win >>5 games in a row. >> >>What is going to be your best estimate for the rating of that program. >>The best estimate is more than 1800 because you can guess that the reason that >>the program performed better is a significant update in the program and not luck >>but the rating that this program is going to get will be lower than 1800 in case >>that it is considered as a stabilized player. > >No, the Elo method doesn't care about program updates. >You may have won 5 games now but lost 10 games yesterday, it isn't clear if you >are above 1800 or not, you must have done something to get a 1600 rating. Even with humans if you see that a stable 1600 player who did not play in tournaments in the last 6 month beat 5 1800 players in a row there is a reason to suspect that the player learned something and it is not luck and this suspect should be used to get better estimate. If you see that the player beats 2600 players instead of 1800 players then your suspect can be changed and you can suspect that the player is cheating by a computer program. > >When you get a challenge you are informed how much you will lose/win pending the >outcome. I agree it should be possible to adjust that increment so if you know >that your program has improved 500 Elo you could get there faster using a higher >coefficent. However, this must break some fundamental fixed mean assumption. Ie. >that in a large pool of players the mean rating should remain constant, if one >goes up another must go down. It is pretty unrealistic that everyone should go >up, but of course with programs that is usually the case. It is also possible with humans that the average of the same team go up because humans learn. I see no reason to assume that the average is the same but if you want to assume it there is no problem. You can reduce rating from everyone and keep the average the same. 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.