Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: how not to calculate performance

Author: James T. Walker

Date: 04:29:26 10/27/04

Go up one level in this thread


On October 26, 2004 at 03:08:09, Uri Blass wrote:

>On October 25, 2004 at 18:21:23, James T. Walker wrote:
>
>>On October 24, 2004 at 17:02:44, Uri Blass wrote:
>>
>>>On October 24, 2004 at 13:04:07, Stephen A. Boak wrote:
>>>
>>>>On October 24, 2004 at 02:12:51, Uri Blass wrote:
>>>>
>>>>>On October 24, 2004 at 01:47:54, Stephen A. Boak wrote:
>>>>>
>>>>>>On October 23, 2004 at 18:47:26, Vincent Lejeune wrote:
>>>>>>
>>>>>>>On October 23, 2004 at 16:37:37, Stephen A. Boak wrote:
>>>>>>>
>>>>>>>>On October 22, 2004 at 18:52:13, Uri Blass wrote:
>>>>>>>>
>>>>>>>>>On October 22, 2004 at 18:30:34, James T. Walker wrote:
>>>>>>>>>
>>>>>>>>>>On October 22, 2004 at 13:32:57, Uri Blass wrote:
>>>>>>>>>>
>>>>>>>>>>>go to the following link
>>>>>>>>>>>
>>>>>>>>>>>http://georgejohn.bcentralhost.com/TCA/perfrate.html
>>>>>>>>>>>
>>>>>>>>>>>enter 1400 for 12 opponents
>>>>>>>>>>>enter 0 for your total score
>>>>>>>>>>>
>>>>>>>>>>>Your performance is 1000 but if you enter 1 to your total score your performance
>>>>>>>>>>>is only 983.
>>>>>>>>>>>
>>>>>>>>>>>It seems that the program in that link assume that when the result is 100% or 0%
>>>>>>>>>>>your performance is 400 elo less that your weakest opponent but when your score
>>>>>>>>>>>is not 100% it has not that limit so they get illogical results.
>>>>>>>>>>>
>>>>>>>>>>>Uri
>>>>>>>>>>
>>>>>>>>>>My take on this is they are using a bad formula or have screwed up the program
>>>>>>>>>>to calculate the Rp.
>>>>>>>>>>The USCF uses Rp=Rc + 400(W-L)/N
>>>>>>>>>
>>>>>>>>>It seems that the USCF does not do it in that way
>>>>>>>>>
>>>>>>>>>They admit that the formula is not correct for players who won all their games
>>>>>>>>>
>>>>>>>>>Note:  In the case of a perfect or zero score the performance rating is
>>>>>>>>>estimated as either 400 points higher or lower, respectively, than the rating of
>>>>>>>>>highest or lowest rated opponent.
>>>>>>>>>
>>>>>>>>>It is probably better to estimate the preformance based on comparison to  the
>>>>>>>>>case that the player did almost perfect score.
>>>>>>>>>
>>>>>>>>>Uri
>>>>>>>>
>>>>>>>>Dear Uri,
>>>>>>>>What is the *correct* formula for a player who has won (or lost) all his games?
>>>>>>>>:)
>>>>>>>>Regards,
>>>>>>>>--Steve
>>>>>>>
>>>>>>>
>>>>>>>For such a player, the error margin = infinity
>>>>>>>
>>>>>>>the perf = average opp +400 to +infinity
>>>>>>
>>>>>>Thanks, Vincent.  I know the formula well.  :)
>>>>>>
>>>>>>I was poking fun at Uri (just teasing) for complaining about 'logic' when in
>>>>>>fact the formula for all wins or all losses is purely arbitrary.
>>>>>>
>>>>>>[I've read that Uri is a mathematician, so I like to occasionally jump in and
>>>>>>comment when he seems to overlook something basic.  All in good fun--I
>>>>>>appreciate his postings and chess programming contributions.]
>>>>>>
>>>>>>I asked Uri what formula would he suggest as 'correct'.
>>>>>
>>>>>I think that it is possible to calculate the performance of a player that get
>>>>>1/2 point instead of 0 point and use the result as an upper bound for the
>>>>>performance of the player that got 0 points.
>>>>>
>>>>>It is not done.
>>>>>
>>>>>Another idea is to assume probability of win draw loss for every difference in
>>>>>rating and to calculate the maximal rating that the probability to get 0 points
>>>>>is 50% or more than it.
>>>>
>>>>Ok, Uri, I accept the challenge.
>>>>
>>>>Assume a player scores 0 points out of 10 total games.
>>>>
>>>>What Win, Draw, Loss probabilities should be used (arbitrary once again, for an
>>>>unknown player's strength, who has scored m=0 out of m games--right?), and for
>>>>what difference in rating?
>>>
>>>The opponents are players with known rating.
>>>
>>>We can calculate the probability of player with rating 1300 to get 0 points
>>>against the players because we have some assumption about probabilities based
>>>on the difference in rating.
>>>
>>>We can choose every different number than 1300 and calculate the probability for
>>>every x so we have a function p(x) and later solve the equation p(x)=0.5.
>>>
>>>>
>>>>What is the 'logic' for your choice of 'difference in rating'?  That choice has
>>>>to be illogical (totally arbitrary), when the new player doesn't yet have a
>>>>rating, right?
>>>>
>>>>ELO does exactly that already--assumes the player is exactly 400 points less
>>>>(i.e. assumes a particular rating difference) than his opponents' average
>>>>rating, against which opponents he has scored m=0 points).
>>>>
>>>>What is a more logical rating estimate for the player than 400 points less than
>>>>the average rating of his opponents?  What is the improved 'logic' for your
>>>>suggested rating approach?
>>>
>>>We want the best estimate for the rating of the player based on the results.
>>>400 elo less than the average of the opponents is llogical because by that logic
>>>if a player lose against 10 players with rating 1400 and 10 players with rating
>>>2400 the player get 1900-400=1500 and it is clear that the player is weaker than
>>>1500.
>>>
>>>Even in the relevant link minimal rating minus  400 and not average rating minus
>>>400 was used but it is still illogical.
>>>
>>>>
>>>>My thesis is that any suggested formula involves as much guessing, as much
>>>>arbitrary choice, as much 'illogical' thinking (because it is totally arbitrary)
>>>>... as the original +/- 400 points rule used in the ELO system.
>>>
>>>No
>>>
>>>You need to give an estimate for the rating.
>>>It is logical to give smaller estimate to player that score 1/2 point and not to
>>>player who scored 0 point.
>>>
>>>Every frmula that does not does not do it simply can be improved to a better
>>>estimate easily.
>>>
>>>If the estimate for players that score 1/2 point is correct the estimate for
>>>players who score 0 point need to be smaller(I do not know how much smaller but
>>>it need to be smaller)
>>>
>>>If you lose against many players the estimate need to be smaller relative to the
>>>case that you lost only against one of them.
>>>
>>>I do not know how much smaller but the formula should promise that it will be
>>>smaller.
>>>
>>>It is possible to test different methods in practise by investigating real cases
>>>of players who got 0 points and find their real level based on games against
>>>weaker opponent.
>>>
>>>The problem what is the best estimate is a practical problem and I did not
>>>investigate it(incestigation can try many possible methods and testing their
>>>errors in predicting future results when you choose the method that reduce the
>>>error to be minimal) but one clear rule is to give better performance for
>>>players who do better results.
>>>
>>>It was not done by the link that I gave.
>>>
>>>Uri
>>
>>Uri it is impossible to come up with a practical score for anyone who is winless
>>or undefeated in his/her first few games.  It does not matter who they play or
>>what the opponents rating is.  The performance formula is used only to establish
>>a "performance" rating in a short match/tournament.  If its the first rated
>>games you have ever played then you are only given a "provisional" rating untill
>>a more accurate rating can be established in later tournaments/matches.  Someone
>>rated 400 points above you has about a 91% chance of beating you and you have
>>only a 9% chance of winning.  If you think you can give an accurate rating to
>>someone who has never won a game in his life and has only played 4 games then
>>you are dreaming.
>
>I did not claim that you can give an accurate rating.
>rating is not accurate.
>I only claimed that you can give a better estimate.
>
>rating is always an estimate and even if you know only the results you can use
>them to calculate some estimate.
>
>
>  Computer programs calculate the performance rating and nobody
>>actually looks at the games to see how strong/weak the player may actually be
>>(In their estimate).  In fact by the time they play their next 4 games they may
>>have improved by 400 points.  I know several people who have done this because
>>they waited 2 years before playing again.
>
>
>It only suggest that the dates of playing should be also used to calculate
>performance and old games should get smaller weight.
>
>  In the case of the one win scenario
>>you mention the link you gave gives two "ratings".  The one you quote is a
>>"performance" rating and the second is the actual "new" rating.  In the problem
>>you quote there is no reason for these two ratings to differ.  They should both
>>show 1067.  Therefore the link has improperly applied the formula which
>>calculated the score below 1000.  IMHO.
>>Jim
>
>
>The estimate for 0 out of 12 against 1400 players should be smaller than the
>estimate for 0.5 out of 12 against 1400 players.
>
>The estimate for 0 out of 12 against 1400 players should be smaller than the
>estimate for 0 out of 4 against 1400 players.
>
>If these rules are not kept then my opinion is that the estimate can be
>improved.
>
>Uri

I still don't follow your logic.  Your assumption that 0/12 is worse than 0/4 is
not logical.  In both cases the actual rating of the person playing may be
something like 500.  In fact it could be the same person that played 4 games one
day and then a week later played another 8 games.  As you say it is only an
estimate untill enough data can be obtained to make a more accurate assesment.
Jim



This page took 0.01 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.