Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: GM Smirin vs 4 comps - Match Predictions

Author: Walter Koroljow

Date: 10:38:38 04/17/02

Go up one level in this thread


On April 17, 2002 at 12:42:12, Miguel A. Ballicora wrote:

>On April 17, 2002 at 08:08:56, Walter Koroljow wrote:
>
>>On April 16, 2002 at 21:13:40, Miguel A. Ballicora wrote:
>>
>>>On April 16, 2002 at 18:34:57, Walter Koroljow wrote:
>>>
>>>>Of course you can combine TPRs of different programs as per common sense.
>>>>
>>>>However, if you wish to be mathematically absolutely correct and precise, all
>>>>you need do is say, "The average of the programs' TPRs is ...".
>>>>
>>>>A little algebra will quickly show that the the average of the programs' TPRs is
>>>>exactly the TPR of the programs taken as a group.  This change of wording gives
>>>
>>>Actually, that is not correct.
>>>
>>>Regards,
>>>Miguel
>>>
>>>
>>I believe it is correct.  Here is a simple example: two programs play a human
>>rated 2700. Program 1 wins and program 2 loses.  then:
>>
>>TPR1 = 2700 + (wins-losses)*400 = 2700+400 = 3100
>>TPR2 = 2700 + (wins-losses)*400 = 2700-400 = 2300
>>Average of two TPRs = (2700+400 + 2700-400)/2 = (2300 + 3100)/2 = 2700.
>>
>>Notice that this is just 2700 + average of 400*(wins-losses).
>>
>>On the other hand, the "team" TPR is:
>>
>>2700 +(wins-losses)*400/2 = 2700 + (1-1)*400 = 2700.
>>
>>This is also just 2700 + average of 400*(wins-losses).
>>
>>This illustrates the general case:  in both cases the (wins-losses)*400 term is
>>averaged over all games.  Hence the same answer.
>>
>>Best regards,
>>
>>Walter
>
>Yes, you are partially right because that is the way that USCF calculate (or at
>least used to be, I do not know whether it changed) the initial rating of a
>player. However, that is an aproximation. IMHO, it is a very gross aproximation.
>It assumes that the players are reasonably close in rating and the elo table is
>linearized in that area of the curve. You can have very distorted results when
>those assumptions are not correct. For instance, the initial rating of Anatoly
>Karpov in USA was ~2500. Probably still is. I am sure is because he played in an
>Amateur US ch (I remember this was 3-4 years ago), beat everybody but could not
>be more than the average of the opponents + 400.
>
>Generally, TPRs are not calculated as USCF does for the initial ratings.
>I have seen that many times that the average of the opponents is taken and the
>calculation is made, using the Elo table, like you play n games with an "average
>player" without introducing that 400 points constant. That is a better
>aproximation. With this type of calculation, your original statement is not
>correct. That is what I meant.
>
>Note that this is still not perfectly correct:
>What would you prefer if you are 2500? playing against two players 2500 and 500
>or playing against two players that are 1500?
>
>The perfect way to do it is to estimate your TPR, calculate how many points you
>expect to get calculating it individually for each opponent and adding it all.
>Then you make sure that this number equals the points that you really obtained,
>otherwise, iterate. Nobody does it in this way for obvious reasons.
>
>Regards,
>Miguel
>

Miguel,

I do not like the TPR calculations as they are done for the same reason you do
not.  For my personal use, I define my TPR as the rating of a player that would
be expected to achieve the result I achieved.  This is, of course, highly
non-linear, a lot of trouble to use, etc.

I believe, however, that the term "TPR", in public discussions, means precisely
the linearized formula I used above.  It is unfair to use a private definition
in such a discussion.  So I used the usual definition in my post above.

By the way, I have done a little analysis on this issue.  My result is that the
linear approximation is very good up to about a difference of 300 points, and
then it breaks down very badly.  So I do not think the difference in definitions
 is significant for the computer-Smirin match (if you are willing to say
computers are between about 2400 and 3000).

I am a little pressed for time, but I will try to dig up my notes and post them.

Best regards,

Walter
>
>
>
>
>>
>>>
>>>
>>>>you an unassailable mathematical position.  And I think the common sense meaning
>>>>is not changed.  If you don't want to bother with this distinction, I won't
>>>>mind. :)
>>>>
>>>>Cheers,
>>>>
>>>>Walter
>>>>
>>>>P.S. Chris - When you say TPR, it would be nice to say whose TPR - Smirin's or
>>>>the programs'.



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.