Re: Re: Re: [computer-go] Statistical Significance

[petrip@xxxxxxxxxxx]

> From:: Heikki Levanto <heikki@xxxxxxxxxxxxxxxxx>
> To: computer-go <computer-go@xxxxxxxxxxxxxxxxx>
>
> Last time I checked (which is many years ago), the Danish Go Association
> used something like the ELO system to calculate Go ratings. They had it

There are several of these attempts. France has one. Also European Gu Federstion has one (GoR)
> Heikki Levanto   "In Murphy We Turst"     heikki (at) lsd (dot)
> dk
> 

Biggest difference between AGA/KGS/IGS style of rating compared to elo is how the adjustment is done.
Elo is simple Linear P-type conroller i.e. you receive/lose points relative to size of error. Error being 
real result vs predicted result Like win was result +0.5 while at equal elo expectation 0.

AGA/KGS/IGS rating make similar aproach error ius measured BUT adjustment step is non-linear.  BAsically idead is Assume players strenghts to ne unknows random variables around the current
strength  estimate. And Find a max likehihood estimate for the all rating being estimated 
based on number of games being observed.

This sort of adjustment can occasionally reduce players strength even when he wins, but that is quite rare. Also it takes into account more than one game in making adjustment.

And it converges faster. Well IGS is slow to converge but as far as I know KGS is pretty fast and still stable.

To put it short P-type controller is no longer state of the art in estimation. HAs not been for ages.

Elo system could easily be modified to include handicaps (fex 1 hcap stone increases chances as much
as 100 Elo points). But Maths behing it should be changed. To PI-controller at least :).


Petri Pitkänen


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