Re: computer-go: Machine Learning and Go

[Markus Enzenberger <markus.enzenberger@xxxxxxxxxxx>]

> As I understand it from Markus Enzenberger's description, I havnt read
> Thore's
>   but
> it seems to be similar
>
> xxxxxxx
> xoooo.x
> xxxxxxx
>
> will have the same representation as
>
> xxxx
> xo.x
> xxxx
>
> is this right?
> If so the some concept(feature) of dead shape should be accounted for.

That's right. And in this special case, it's clearly a weakness.
Although the loss of information about the shape of a string is
a good thing for live strings, it is bad for dead strings, if the
strings are of intermediate size and if it is the only possibbility for
the surrounding strings to get 2 eyes by capturing it.

But this is a very special case, occuring only rarely in games at the
level where Go programs play at present. Much more often, big groups
are killed, providing plenty of eyespace for the opponent.
The benefits of representing strings as a single node clearly outweighted
the disadvatages in my networks.

> If so the some concept(feature) of dead shape should be accounted for.
 
Yes. And it is simple. Add an attribute (or input value for neural
nets) to the string node, telling whether it is worth 1, 1.5 or 2 eyes for the
opoonent if captured. This can be looked up in an eyeshape library.

- Markus
                                                                               
-- 
Markus Enzenberger | http://home.t-online.de/home/markus.enzenberger