[blml] Swiss Teams formats

[Jean-Pierre Rocafort]

Steve Willner a écrit :
>> From: Jean-Pierre Rocafort <jean-pierre.rocafort at meteo.fr>
>> 2-day competition, 14-30 teams, 80-90 deals, 5-8 rounds, 1 or 2
>> predefined rounds based on seedings. the particularity is the use of a
>> bonus system to prevent swissing and to make up for the disparity of
>> opposed teams: complementary VP are added to each team at the end of
>> each round (except 1st and 2nd) according to their present ranking (0 to
>> the last team, the most for the leader). the exact amount of bonus is an
>> arbitrary parameter, the tuning of which adds to the debate.
> 
> This is perhaps a little better than doing nothing, but it seems far 
> worse than matching according to record (possibly with a bit of weight 
> given to seeding as in accelerated Swiss), then adding in the "strength 
> of schedule" correction at the end, when you know how all the teams have 
> done.

i am not certain to understand what you mean by "record". with, for 
instance, 16 teams seeded 1 to 16 according to masterpoints, it works 
following:
1st round:
1vs2, 3-4, 5-6, ... 15-16

2nd round:
1-15, 2-16, 3-13, 4-14...

next rounds: according to rankings, 1-2, 3-4... excluding rematches
bonus added from the end of 3rd round: 0 to 16th ranking, k to 15th, 2k 
to 14th,... 15k to 1st.


>  The exact amount to add in isn't clear, but simulations should 
> tell us what it should be.
simulations and analyses are needed to evaluate the optimal value for k 
but conclusions are not obvious. it is sensitive to the number of teams, 
number of rounds, homogeneousness of field. practical values currently 
used for parameter k are in the range 0.1 to 0.3.
> 
>>   one way to analyze the accuracy of a particular swiss teams result is
>> to compare it to an "objective" modelisation of the performance of the
>> competing teams. this modelisation is an extrapolation of the swiss to
>> what would have been the results in a complete round robin between all
>> teams, using the methodology of missing-data reconstitution: results
>> of actually played matches are retained; for non-played match, 
>> artificial results are computed, statistically the more compatible with 
>> other results of the 2 teams involved. (technical details available upon 
>> request)
> 
> If I understand what you mean, I long ago suggested something similar. 
> I'd do it directly by solving appropriate "normal equations," given the 
> actual results.  The biggest problem with this approach is that it is 
> entirely opaque to the players, but it is theoretically the best method 
> because it takes full account of all results.
the intent is not to use the method for the final result but only as a 
mean to evaluate other (non opaque) methods by comparison. 
notwithstanding, whatever method used is judged insane and random by all 
players, especially when they don't win.

> (Actually there are 
> several possible methods, not a single "best" one, in this family.) 

the method is optimal in the sense of "lesser squares". it's a very 
elementary result in statistics theory.

jpr

> I've suggested the same approach to player ratings, where it might 
> receive more support because nobody expects ratings to be easily 
> computable by players themselves.
> 


-- 
_______________________________________________
Jean-Pierre Rocafort
METEO-FRANCE
DSI/CM
42 Avenue Gaspard Coriolis
31057 Toulouse CEDEX
Tph: 05 61 07 81 02      (33 5 61 07 81 02)
Fax: 05 61 07 81 09      (33 5 61 07 81 09)
e-mail: jean-pierre.rocafort at meteo.fr

Serveur WWW METEO-France: http://www.meteo.fr
_______________________________________________