Author: Robert Hyatt
Date: 10:35:46 03/17/99
Go up one level in this thread
On March 17, 1999 at 09:58:04, Andrew Dados wrote: > >On March 17, 1999 at 08:56:49, Albert Silver wrote: > >>This whole story about Mark's account screwing up the ratings on Chess4u has >>been somewhat interesting. No doubt a few will disagree. The reason is that NO >>ONE except for Hyatt, though for different reasons, actually gave any credence >>to this. Chess4u is right, but not about Mark. The accounts that inevitably >>cause inflation are the ones that use more than one program or accounts where a >>lot of testing is done. Suppose I have, as Mark did, Hiarcs 7 running on a >>PII-450 and it gets an official rating of 2800. No problem as it is indeed >>playing at that level and it's results correspond accordingly. Now suppose after >>about 2 months, I see the latest version of GNU chess out. The author claims it >>is vastly improved and should be playing much better, though no one knows just >>how much. I decide to test it with my account. GNU chess is not a 2800 player, >>but when testing starts it is playing with a 2800 rating. It gets trounced by >>the super opposition and the rating drops until it stabilizes at around 2300. I >>am not personally worried as after the testing is done, H7 will obviously regain >>it's lost points. The problem is that 500 points were spread out in the pool and >>they don't properly represent an increase in strength on the opponents' part. >>When I get back, I don't go to 2800, but a bit higher as I am now playing the >>same opponents, but with slightly higher ratings. If a program undergoes >>testing, and experiences severe rating fluctuations while it is being tested, >>then the same phenomenon takes place. Bob is obviously already aware of this as >>his notes to his Crafty account on ICC state that opponents who clearly play him >>ONLY when Crafty's rating is high but never when it is at a low, will be >>'noplayed'. >> >> Albert Silver > > Some loose remarks on a subject: > According to your post *any* bigger rating fluctuation cause ratings >inflation. Same holds true when I come to server tired or, like some weekend >players, drunk. How does it differ from your point, dunno. Other rating - >related problem I noticed is deflational Glicko rating system implemented on >fics, which does not preserve points pool.... but that's a different story >(Well.. one guy blames computers for there is less and less players over 2000 >and people seem to make little progress over time). > I've met some people who claim that computer accounts are overrated; a few >who claim that they skew rating system and many, who just think there are too >many (C)... >....yeah... blame the puters... :) > :Everybody, even aware that it makes little sense, want to compare rating >numbers from different systems (like icc and uscf). When they are lower rated on >chess servers (rarely) >they blame 'system' or (C)... those numbers are miningless as absolute values >and yet '2000' is way better then '1950'.....not to mention that magic 2200.. > >-Regards- >Andrew this is another interesting aspect. The "K-factor" in the rating calculation was arrived at most likely by 'trial and error'. IE it was tuned for tournament chess where you might play 5-10 games per month, and K has to be large enough to let your rating swing as much as it should. But what if you play 100 games per day. I'd almost think K should shrink _way_ down... ie as in the "glicko" rating system that is _much_ more stable. IE what is the probability that Crafty is actually 200 points worse than yesterday, when _no_ changes have been made? Very low. Yet if I play an equal-rated player, my rating changes by +/- 16 after one game. For a server, I believe this is _wrong_.
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.