Yamato probably meant at the frontier between random playouts and tree
search. I do something like that in Crazy Stone, too. For instance, if
the last move in the tree was an atari, and a quick tactical analysis
shows that the opponent cannot survive by extending, then the first
random move of the playout should not extend.

Rémi


Thank you for that clarification. I'm sure you're right.
I don't fully agree with Yamato though.
Semeais foul up the evaluation in two distinctly different ways.
The obvious way is that a playout gets the wrong result for a semeai.
To avoid that, "cleaning up" semeais at the border between tree and linear
playout, as Yamato suggests (if I interpret "do something" correctly),
should be effective.
After clean up, the random playout certainly has a better chance of getting
the correct result.
But there is a more insidious, and in my opinion more devastating effect in
the tree part of the search:
The losing side actively seeks out the variatons that muddy up the
evaluation.
Frolicking in the permutations of the possibly multiple semeais on the
board, the would be losing side manages to hide the true result "behind the
horizon".
Eliminating, or atleast reducing the frequency of, this obfuscation through
trashing seems a promising approach to further improvement.

Stefan

_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go

Reply via email to