terry mcintyre wrote:
>> On Thu, 2008-09-04 at 18:07 +0200, Rémi Coulom wrote:
>>> When the playouts evaluate a critical semeai the wrong way, then no
>>> supercomputer can help, even at long time control. Semeais require a
>>> better algorithm, because no computing power can search them out with
>>> a
>>> tree, and playouts have to be extremely intelligent in order to
>>> evaluate
>>> them correctly.
>
> Has anyone tried implementing the ideas in Richard Hunter's "Counting
Liberties" in a manner which can be used to guide the playouts?
>
Static analysis is hopeless. Local search may work with some level of
precision but you will have to read much further ahead than in Richard
Hunter's model positions before you can even reliably recognize them.
In particular I wouldn't advise trying to cancel outside liberties
without playing out the moves in some order, otherwise detecting the
need for approach moves is extremely difficult.
GNU Go makes an attempt at reading semeais but has trouble already
differentiating between outside liberties, inside liberties, and
eyespace, not to speak about weird interactions with the tactical
reading (which mostly doesn't understand semeai) in low liberty
situations.
If you think this should be simple, have a look at e.g. page 53,
diagram 17, in Hunter's book or try to do a static evaluation of the
semeai in upper left corner at
http://trac.gnugo.org/regression/regress.plx?nngs2:150
Things to think about for the latter position:
* How big eye does black have around F19?
* What can white do around E9?
* How to handle white sacrificing the C11 tail?
* Is there a liberty to be gained by playing A14?
* Does white need an approach move at M18?
* Does black need an approach move at B10?
* What is A16? Eyespace, some other kind of liberty, or neither?
For some more semeai fun I can recommend
http://trac.gnugo.org/regression/regress.plx?semeai:133
(which has an incorrect solution in GNU Go's test suite).
/Gunnar
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
