[Computer-go] issues on 19x19
darren at dcook.org
Thu Apr 28 19:39:45 PDT 2011
>> That's a great question. I sort of assumed that one of the improvements
>> Remi worked on since old CrazyStone days was adding RAVE. But maybe I
>> assumed wrong...?
> No, as far as I know, Crazy Stone is not using RAVE. But I think Crazy
> Stone will be stronger if Remi combines RAVE with progressive widening,
> like some strong programs (pachi, Aya, Gomorra...) did.
What makes you think that? (real question, not rhetoric)
RAVE is a bit of a crude heuristic: if a move is good deeper in the tree
it is good at the top of the tree. It works for the same reason monte
carlo playouts work: in go, most of the time most of the positions on
the board are not that sensitive to move order. (*)
Do people think RAVE is just a stepping stone and eventually it will
have to be replaced by something that understands better what is going
on. Or do you think when we finally roll out the "9p bot" that RAVE will
still be in there?
By the way, is "bonobot" stronger than Aya on equivalent hardware, or
are they about the same?
*: Despite the know-it-all tone of that paragraph, I still don't really
feel I've fully "got" either MCTS or RAVE. Insights of other people are
Darren Cook, Software Researcher/Developer
http://dcook.org/work/ (About me and my work)
http://dcook.org/blogs.html (My blogs and articles)
More information about the Computer-go