[computer-go] evaluation function

[stephane willemin <stephane_willemin@xxxxxxxx>]

Hello all,

I am trying to make a go engine, and wanted to have your opinion about some thoughts i had about computer go programing, and also ask some questions.

I think on large boards, a go program has to exploit somehow the locality of the game to counter the high branching factor, and do local search when it is possible.This is what is done by current programs wich use sub goals about tsumego and block tactics in their evalution function. But i think this is not the best way to use locality, because often knowing whether or not a block can be captured is not enough, you might actualy ruin your position by trying to defend it, even when you succeed. Of course i dont know of a better way, but has anyone thought about it and can share some insights?

I was also wondering if highly selective search coupled with a good evaluation function (no local search) could be enough to make a strong go program on 9*9 (where locality is less a problem)

What do you think?

Finaly, i would like to address the problem of designing an evaluation function (i mean a static evaluation function)

This is harder than it seems to have a program just recognise obvious properties like secure territory and clearly dead groups. Has anyone tried to build a static evaluation function? Can you share about this?

---

Appel audio GRATUIT partout dans le monde avec le nouveau Yahoo! Messenger
 
Téléchargez le ici ! 
 
_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://www.computer-go.org/mailman/listinfo/computer-go/