Re: [computer-go] Changing the subject

["Jeffrey Rainy" <jrainy@xxxxxxxxxxxx>]

> > Now I'm more interested in what the algorithm is able to find than in
> > the performance of the function as a Go player.
>
> How do you define the performance of an evaluation function if not by
> how well a program using it plays?  Or am I confused about something?

I use its predictive power. From a given position and the move made by a
high-ranked player, I rank all the possible moves acoording to the
evaluation of the resulting position. Ideally the move played by the
high-ranked player should be one with high evalution. So, the percentage of
moves that were ranked lower than the played move is the predictive power.

Simple stuff like: "play near last move" gives you about .8 predictive
power. Complex stuff like my NN evaluation function also give around .8, but
without needing to know what the previous move was (which is either
interesting or useless depending on you POV)

J.

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