Author: Peter Fendrich
Date: 02:05:04 08/03/98
Go up one level in this thread
On August 02, 1998 at 08:01:28, Robert Hyatt wrote: >On July 31, 1998 at 20:51:56, Bruce Moreland wrote: > >> >>On July 31, 1998 at 14:00:04, Peter Fendrich wrote: >> >>>The ratings on ICC and FICS are really swinging. IMHO they are using a >>>completely wrong approach to handle this. The ELO formula is not at all suited >>>for the kind of events taking place on these real time severs. The original ELO >>>formula is using a kind of constant a 'dampening' factor for varitaions during >>>time. At least on ICC they have tried to use another, more dynamic but similar >>>method to handle this problem. There are new, much better, ways to deal with >>>this. Nowerdays used within some of the space, aero and automotive applications. >> >>Both ICC and FICS tried to use the Glicko system. >> >>On ICC, they made it an "extra" rating system, but I don't think anyone paid >>much attention to this, so it is removed from the "finger" notes. It is still >>there, but to see it you have to use "yfinger". >> >>On FICS, they made this the only rating system. I am not there much, but I >>remember seeing a lot of posts about this, people were upset about this system >>because if you played a lot, your rating tended to stay constant no matter how >>you did, and people wanted to have more movement in their ratings. So I believe >>that on FICS they patched Glicko somehow, so that ratings would still move a >>bit. >> >>bruce > > >the problem with "Elo" is that the "K" factor was statistically derived from >the typical number of rated games a person would reasonably play in a year, and >the max expected rating change of a person over that time frame. It is totally >wrong for a server where (say) a program plays 20,000 games per year. Because >we see the huge swings that result from this. Chances are that if you are a >2,000 player today, you will be a 2,000 player in 6 months, regardless of how >many games you play, So it would be difficult to pick a formula that is fair >to those playing a dozen games a year and to those playing thousands. What ELO tried to cover with the K-factor was the time and that was doable because it was impossible to play a lot of rating games in a short period. A better parameter on the servers is the age of games. The older they are the less importance they have. Another problem with the servers is the variation of opponents. If you play 50 games against 50 different opponents and the another 100 games against only one opponent the last 100 games will affect rating more than the 50 first games. This is incorrect and will give a bias to the rating and furthermore gives possibilities to frauds by having two different id's and beating the hell out of one of them. By beating yourself in a lot of games it is possible to raise your own rating... A good formula should have restrictions in variation of opponents. The third problem is how to set the start rating for new id's. It's hard to find a sound way to do this. In a pool of players where people are registered and unregistered a lot during time, the rating for the rest of the group will be biased (compared to other rating pools). Every new player gets new rating that adds on to the rating pools total and for every unregistered player some rating points are disappearing from the total group. For example I could register and (hopefully) win a lot of rating points from the others. After a while when I quit all the rating points that I won will be gone and impossible to "win back". Because people are changing their strength during time there is no correct formula to deal with the absolute rating point at every specific time. The rating will always be some kind of weighted average over time. There are better and more dynamic ways to handle this situation than the ELO formula and its relatives does. Especially for chess severs and the blitz games played there. One could still use the same Bell curve but apply idividually set "filters" when computing the ratings. I will work on this when time is given... //Peter
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.