Re: [computer-go] 9x9 Ratings

["William M. Shubert" <wms@xxxxxxxxxx>]

What you have here is something that is pretty a maximum likelyhood
system, but with the probability equation set to say that the stronger
player has a 100% chance of winning and the weaker player has a 0%
chance of winning.

Since you have come so close and all you need to do is plug in a "real"
probability of win equation, why not just take one and do so? You can
take the one from KGS if you want (see it at
http://kgs.kiseido.com/en_US/help/math.html) or you can dig up mlrate,
which NNGS uses, which uses a slightly different equation.

Since all the math will be done by a computer, doing a real maximum
likelyhood will not be much harder than what you have already.

On Tue, 2005-03-15 at 23:24 +0100, Christian Nilsson wrote:
> I gave my rating code a bit more thought and realized that it was
> quite flawed. Since our programs most likely wont improve after each
> game, there is no reason to update the rating after each game using
> the old rating and some other information. I therefore tried an
> alternative. You may skip this section if you don't care.
> 
> I store all results in a vector where each index corresponds to an
> opponent rating, 9d to 30k. The results are spit into wins and losses
> for each rating...I then proceed to try out all possible ratings, 9d
> to 30k, and calculate a "badness" value. All games that are lost to a
> weaker opponent will be added to this badness value along with all
> games won against a stronger opponent. I then simply pick the rating
> with the least badness. I tried it for a number of players on kgs (
> 19x19 games mainly ) and it matched the kgs-rating for all except some
> 9d's. This lead me to believe it may very well be a better estimate
> than the previous.


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