Re: Monte Carlo computer go methods


Robert Maas, see http://tinyurl.com/uh3t a écrit :
[ ... ]
On 11*11 you see your opponent start to stake out territory loosely
some of the time instead of always going for a secure corner group.

Denis Feldmann <feldmann.denis.asupprimer@xxxxxxx> wrote:
And how is this an answer to "Could you find any of this by yourself"
(not to mention that this is *certainly* not what you find by watching
MC computers now, and there is no reason to believe it will change :
the most efficient MC play looks always strange, as he just try to
maximize his probability of win, and not its score...


Agree with Denis: MC can find many good moves yet there
could be certain good moves MC is unlikely to find. It applies its
own lookahead function, so ping-pongs within its own vacuum.
Given sufficient time MC will eventually find a good move it was
not finding previously, though perhaps not rapidly enough to stay
within tournament time limits. MC could benefit by heuristics that
compile positional databases, maybe combination with MoyoGo.
There are "sharp" positions where MC might fail to notice unusual
continuations. There are also deliberate "ko" maneuvers where
MC loses the thread of "reading" capability. If the program only
tries to maximize "win probability" then it won't obtain advantages
of "score difference" which also buffer against "loss probability."



- regards
- jb

-------------------------------------------------------------
"KLA sold Kosovo Serb organs in Albania"
http://www.wvwnews.net/story.php?id=3899
-------------------------------------------------------------

.