Author: Robert Hyatt
Date: 18:48:44 01/15/03
Go up one level in this thread
On January 15, 2003 at 16:43:16, Uri Blass wrote: >On January 15, 2003 at 13:00:32, Robert Hyatt wrote: > >>On January 15, 2003 at 12:20:14, Uri Blass wrote: >> >>>On January 15, 2003 at 11:51:53, Robert Hyatt wrote: >>> >>>>On January 15, 2003 at 11:09:32, Miguel A. Ballicora wrote: >>>> >>>>>On January 14, 2003 at 20:47:46, Robert Hyatt wrote: >>>>> >>>>>>On January 14, 2003 at 19:47:21, Uri Blass wrote: >>>>> >>>>>[snip] >>>>> >>>>>>>I do not suggest exactly how to do it and it seems that the problem does not >>>>>>>interest the ICC. >>>>>> >>>>>> >>>>>>ICC is not interested. FIDE is not interested. USCF is not interested. In >>>>>>fact, _no_ chess federation I know of does the initial rating differently than >>>>>>what is done today. >>>>> >>>>> >>>>>I told you already, FIDE does not do it in the same way. >>>> >>>>Have they changed this recently? My last official rules from FIDE used >>>>the _same_ "provisional rating" formula that everyone else uses. The >>>>classic win+400 + draw + lose-400 / N formula that ICC uses. >>>> >>>>So I guess I don't understand what you mean, assuming they _have_ changed >>>>the way that they calculate _initial_ ratings. >>>> >>>> >>>>> >>>>> >>>>>>You are taking a rating as an absolute value. It is _not_. It is an >>>>>>estimate of how you would do against the group (pool) of players you compete >>>>>>in. The current TPR approach is _exact_ in that regard. Even if it has >>>>>>nothing to do with how you would do against other players. I have _yet_ to >>>>>>see anyone suggest an alternative. Just complaints about how it is done now. >>>>>> >>>>>>Without suggestions on a better way, complaints are not very useful... >>>>>> >>>>>>"I don't like that, fix it" is _not_ going to produce changes. >>>>> >>>>> >>>>>Please read my other message. First, you posted "I am waiting a real >>>>>suggestion". I did (in two messages), you ignored them. Uri mention another >>>>>possibility. >>>>> >>>>> >>>>>>>If ICC cares about creating a better rating system >>>>>>>They can give give 10000$ for the people who find the best rating system. >>>>>> >>>>>>I believe Elo did that a _long_ time back. It has certainly stood the >>>>>>"test of time". >>>>> >>>>>The theory yes, but the implementation is far from perfect and many people >>>>>suggested improvements, including Glickman and K. Thompson. >>>> >>>>You realize why the glickman system was developed? The Elo system had >>>>some basic assumptions, one of which was that active players might play one >>>>event per month. That means their rating should change more quickly than >>>>what is happening today where on a chess server, a person can play 20 games >>>>in a single day. The Glickman formula slows your rating change as you play >>>>more games over a short period of time. IE on ICC Crafty's rating can change >>>>+/-200 in a single afternoon as a loss to a player rated 300-400 points lower >>>>will cost it 32 rating points in one game. In the glicko system, that 32 points >>>>would be more like 1-2 points if it had been playing lots of games every day. >>>> >>>>Elo wasn't wrong. But his assumptions are not always reasonable in a >>>>world connected via the internet. But for _either_ rating system, the >>>>initial rating is calculated the same way because no better way has been >>>>found. >>>> >>>> >>>> >>>>> >>>>>>>It seems that they do not care so they will not do it. >>>>>>> >>>>>>>I have also definition that we can compare based on it different rating systems. >>>>>>>A rating system should give the expected result in every game. >>>>>>>It is easy to use the sum of squares to find the practical error and the rating >>>>>>>system that gives the smallest error is the most logical rating to choose. >>>>>>> >>>>>>>Uri >>>>>> >>>>>>Can you spell "Elo"??? >>>>> >>>>>This is rather insulting. You are dealing with Uri and me like we are two >>>>>schoolboys and you are not more qualified than any of us in this matter. I do >>>>>not mind it if it comes with some content, but you are not even reading the >>>>>messages. Uri pointed out a problem and you come with the >>>>>"'I don't like that, fix it' is _not_ going to produce changes." >>>>>What? it is not possible to post anymore something about something that you do >>>>>not like? Before you find a solution, you have to identify a problem and discuss >>>>>about it. Not to mention, suggestions were provided. >>>>> >>>>>Miguel >>>> >>>> >>>>It wasn't intended to be insulting. It was intended to point out that >>>>what you are talking about is _exactly_ what the Elo system was designed to >>>>do. >>>> >>>>Neither of you have produced any _reasonable_ approach to initially compute >>>>a players rating. That seems to be all that is in debate here, as once the >>>>provisional period is up, the Elo system seems to be accepted by all, even >>>>with the minor "flaw" in the assumption of what K should be. But for the >>>>provisional period I have seen _no_ suggestion that makes any sense yet. I >>>>have asked you in another post to simply make a precise formulation of an >>>>algorithm that you think is better. I can then show you why it is not better >>>>the current simple TPR average that everyone I can find is using, FIDE >>>>included according to a web search a few minutes ago. >>> >>>Here is a simple algorithm that is better and I think it was suggested by >>>migual. >>> >>>1)If the player has only wins or only losses do not give the player rating. >> >>Bad idea. Strong player at a university could win every tournament game >>he plays there. He gets _no_ rating??? > >no rating means that we do not know the rating and not that the player is weak. > >I agree that no information is not a good idea. >I suggest that the player will get information but the information will include >only lower bound for rating. > >lower bound can be calculated by reducing 1/2 point from the result and >calculating rating. >> >> >>>2)If the player has not only wins or only losses decide to do the following >>>steps(note that this is not the fastest way but I care only to make the way >>>clear and not to do it fast). >>> >>>a)Guess a rating that for the player based on results >>>b)calculate for every game the expected result based on the rating(the elo gives >>>expected result for every difference of rating) >>>c)You will find that the result is too low or too high >>>d)if the result is too high reduce the rating by 1/2 and repeat the process >>>until you find that the rating is too low. >>>e)if the rating is too low increase the rating by 1/2 and repeat the process >>>until you find that the rating is too high. >>>f)You will get 2 numbers with difference 1/2 and the rating is between them >> >> >>and you will get a number that is _very_ close to the TPR number if you measure >>it over 20 games. That is the point.. > >The number that I get cannot be reduced by winning. >The order of games is not important. And that is a problem for estimating a rating. I beat a 2000 player. I then beat 10 1000 players. My rating can't drop below what I got for beating that 2000 player? That's not reasonable. My rating _could_ be based on a simple lucky game? > >Note that my opinion is that this is not the best idea because I think that the >order of games should be important because players may learn and improve from >playing so the last games should be more important. I disagree if the games are played in a single event. If 1/2 are played this year, 1/2 next year, then I do agree with you. But over a single event, a player's strength is not going to change significantly. > >I do not know how much more important but they should be more important. >It is espacially the case when computers are involved and there may be an >upgrade in the hardware or the software. Same thing happens with humans. Someone explains "distant majority" and "distant passed pawns" and that player's endgame suddenly takes a jump up. > >Uri > >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.