On 12/30/2010 02:48 AM, Robert Jasiek wrote:
Nick has said that a 2007 Hungarian RAVE paper was the theoretical
breakthrough. Is this its URL?
http://zaphod.aml.sztaki.hu/papers/ecml06.pdf
It's very hard to say exactly which paper was a definitive breakthrough.
They all build on other work, going all the way back into the 1990s.
The paper that you link to is the first description of UCT, but it
wasn't used for Go in the paper. You can find a copy at archive.org:
http://web.archive.org/web/*/http://zaphod.aml.sztaki.hu/papers/ecml06.pdf
The MoGo team applied UCT to Go with great success, using the idea of
building incremental trees from CrazyStone. Another big factor in their
success was the hand-picked patterns and the idea of using sequences (if
a pattern matches near the last move, play it; this gives plausible
sequences). The landmark paper that describes these techniques, and that
many programs used as a blueprint was the first MoGo paper: "Modification
of UCT with Patterns in Monte-Carlo Go",
http://hal.inria.fr/inria-00117266 .
The MoGo team kept on advancing rapidly and introduced RAVE as their
next big improvement. Another big improvement is the use of prior
knowledge to seed the starting winning percentage for nodes in the tree.
There are other improvements, but I'm not going to list everything here.
Read the papers about MoGo if you want to know more.
The two strongest programs since MoGo are Zen and Many Faces. My
understanding is that they have mainly focused on improving Go knowledge
into the evaluation function.
Congratulations also to all the theorists! Without their great
discoveries, programs would still be weak. Might somebody please give an
overview on the relevant theories and how they work?
I'd start by reading the Sensei's page: http://senseis.xmp.net/?MonteCarlo
You can read the MoGo papers and trace the references if you want a
detailed understanding. You can also search this list for many past
discussions of these topics.
One thing keeps bothering me though: What does all the strength
improvement give us humans for better understanding the game strategy?
Not much, but eventually a strong computer program can be used to test
theories in a more automated fashion, so that's something.
_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
