[computer-go] tactical search in Many Faces
Mark Boon
tesujisoftware at gmail.com
Thu Sep 28 10:28:11 PDT 2006
On 28-sep-06, at 13:51, David Fotland wrote:
> I read each string with fewer than 4 liberties several times, to
> distinguish between four possible states. All of these searches
> are done during evaluation, before finding eyes, groups,
> life.death, and territory.
>
> 1) String is dead. Prey moves first, prey wins all kos, hunter
> answers all threats to adjacent strings, and prey is still
> captured. These strings can be eyes or connections for the attacker.
>
> 2) String is safe. hunter moves first, hunter wins all kos, prey
> answers all threats to adjacent strings, and prey can still make 4
> liberties or 2 eyes always. These strings can be part of living
> groups.
>
> 3) String is threatened. String is not dead (from earlier
> search). Hunter moves first, prey wins all kos, hunter answers
> threats to adjacent strings, and string is still captured. These
> strings can be unconditionally captured when hunter goes first, so
> they are half-eyes.
>
> 4) String safety is unclear. The result is a ko, or the search was
> too big to resolve one of the above choices. This string might be
> captured, so it might be part of an eye.
>
> I limit each search to 150 unforced nodes. I track which board
> locations affect each search, so the evaluation doesn't repeat
> unnecessary searches. I have pretty good move ordering, so the
> search is a simple alpha-beta without iterative deepening, but with
> a transposition table. An unforced node is a position where the
> move generator gives more than one move. In a ladder, the move
> generator usually only generates one move at each ply, so a ladder
> of any length counts as zero unforced nodes.
>
> David
>
>
David,
"hunter answers all threats to adjacent strings"
sounds like you'll never sacrifice stones, which I'm sure can't be
the case.
Mark
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://computer-go.org/pipermail/computer-go/attachments/20060928/17f0501d/attachment.htm
More information about the computer-go
mailing list