Re: [computer-go] Evaluation function / move ordering for gnugo

["Jeffrey Rainy" <jrainy@xxxxxxxxxxxx>]

Hi,

From: "Don Dailey" <drd@xxxxxxxxxxxxxxxxx>
>Jeffrey, Can you describe some of the features of your evaluation?  We
>are  always interested  in exploring  ideas.  One  question I  have is
>whether it has  some sense of dead groups or is  that to be determined
>dynamically (by search and/or patterns?)

From: "Dave Dyer" <ddyer@xxxxxxxxxxxxxxxxx>
>I once spent a lot of time with this technique, and do not recommend
>it.  For example,  you'll get nearly as good a result if you rank
>possible moves strictly by the distance from the previous move.

I'm in the process of writing up more clearly what my evaluation function
does. It is using neural nets and the following features :
- string color
- string size
- string liberties
- hypothetical string liberties, given 1 extra move
- influence (empty area nearest to each string )
- all connections between string and between string and border
- (almost done) Benson's algorithm information like eyes and
small-x-adjacent regions
* it is not using any information about previous move, though.

The simlpe "play close to previous move" algorithm beats it, in term of
predictive power. However, the two are *not* very correlated. The
correlation coefficient is 0.18 ( on some 10 random games ). Also, my
algorithm has a lower standard deviation.

So, I'll combine both, to see how good a move predictor it makes. I'll post
more detail about the inner working of the evaluation function at some
point, but I need to write it up clearly.

Thanks.
Jeffrey.

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