[computer-go] Caching of local search in NeuroGo?

[Mark Boon]

On 10-aug-2006, at 8:00, Rémi Coulom wrote:

> Yes, you are right. In fact, I also count the liberties obtained by  
> connecting to frindly strings in my code. That's why I added the  
> other comment in my message. I did not remember everything very  
> well because I wrote that code long ago. Sorry for the confusion.
>
> I am pretty sure I do not miss any good move, because I tested my  
> code on a huge number of randomly generated positions.
>

I see, you look at liberties of friendly neighbours as well. That  
can't work in all cases either. For example:

OOOO.O
OXXX..
OXOXXO
OXOOXO
OXX.XO
OOOOOO

Can X be caught as a 2-liberty problem or not? I say it can, the way  
you described it to me it can't

So now you need to account for possible liberties by capture as well.  
And probably there are other kinds of counter-examples, there always  
are. It's not clear to me you'd gain much time that way either, if  
any at all. The extra work needed to check the neighbouring points at  
every move versus reading an extra move occasionally. Note that your  
heuristic saves at most one move at this kind of branch, because the  
criteria must be such that it's immediately clear the stone escaped  
after playing there. If it wasn't clear after one move then I  
question the accuracy of the heuristic even more than I already do.

Mark



-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://computer-go.org/pipermail/computer-go/attachments/20060810/7ea7e6ca/attachment.htm