Re: Re: [computer-go] Statistical Significance

["chrilly" <chrilly@xxxxxxxx>]

> Elo ranking is quite slow to converge. Furthermore is has tendency
deflate. That is why is not used in Go general.
>
>As measuring strenght difference between two player it is overkill. For
problem at hand simple set of games will do. It is just question of how many
games one needs.
>
>Any Go-server maximum likelihood rating system probably beats ELO both in
accuracy and convergence speed.
>
>Petri Pitkänen
>
There is general consensus  in chess to use the Elo system. It is not
perfect, but nobody has porposed so far a convincing alternative.
The slow convergence is a feature and not a bug. There is the - not
unreasonable - assumption that someones playing strength is relative
constant. Loosing a few games due to bad circumstances or simply bad luck
should not influence the rating too much.

What is - concerning the rating system - the difference between Go and
chess? What assumptions of the Elo system are not valid in Go?

Note: One know problem in chess is, that the distribution of the result is
in the tails not a normal distribution as assumed by the system. The
Elo-System underestimates the score of a considerable weaker player. This
has to do with the great draw-margin in chess. One consequence is, that e.g.
Kasparov does not play against players below 2600-Elo (He has 2800). He
would loose Elo by playing against such opponents.
Computers have nowadays the same problem. If one wants to surpass Kasparov
on the Elo-scale, one can only play against very strong human opponents
(Officially there is no Elo-Rating for programms).

Chrilly


_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://www.computer-go.org/mailman/listinfo/computer-go/