I watched the video and my question is if Erica uses dynamic Komi? My
impression was that it early on chose variations that simplified the
game. Then there was a huge fight with a lot of strong moves from both
players, and in the end of the fight it was an even game when the
endgame started. I had the feeling that Erica played as if it had the
game under control all the time but then in the end it crumbled. At
This is something I have seen often. Perhaps dynamic komi would be
crucial in handicap games against pros? By playing more aggressive in
the opening the program may more often enter the endgame with a clear
advantage, or even win a knockout victory.
My thinking here is that the main weakness of MCTS on 19x19 in general
is to handle multiple parallel fights. When a big life and death fight
happens the effective branching factor is small and the program can
read out the fight correctly. But as soon as the position is quiet
with a lot of large moves left the effective branching factor goes up
with 4-5 local fights of more each with several possible candidate
moves.
I guess the top programs rely on some kind of pattern matching for
move selection. I would expect pattern matching be most effective
early and late in the game. Early because situations tend to be
standard and an open center make strategy similar between games so
that pattern generalize well to new position. In the late endgame
patterns should be effective because the endgame becomes more and more
local and tactics more and more simple as groups become strong.
In the early endgame however correct play depend on many stones placed
in patterns that will not match any patterns in a database because
there are simply practically an infinite amount of such patterns.
Small patterns will of course match but they will generalize very
poorly to new situation at this stage.
So a good pattern database to set priors in the search tree may become
ineffective in the late middlegame and early endgame making the
problems of multiple parallel fights even worse. And this is what the
young pro felt.
-Magnus
Quoting steve uurtamo <[email protected]>:
it'd definitely be interesting to know what move (or small region of moves)
she's referring to.
s.
On Sat, Oct 2, 2010 at 11:12 AM, Olivier Teytaud
<[email protected]>wrote:
My two cents: Li Yue told me in the past that a main weakness of computers
was in the "big yose".
Computers are (I think) strong in small yose, but weak in big yose.
I don't know if Rina Fujisawa refers to this by "middlegame".
Olivier
who has a book that extends to the middlegame?
;)
s.
On Sat, Oct 2, 2010 at 8:39 AM, Hiroshi Yamashita <[email protected]>wrote:
You can see Japanese biggest newspaper news. (In Japanese)
http://www.yomiuri.co.jp/national/culture/news/20101002-OYT1T00668.htm
--------------------------
Translation:
Strongest program played self-destruct. 12 year's old professional
defeated computer.
Professinal Go player FUJISAWA Rina won against Taiwan's program Erica
with 6 handicaps game.
She bacame pro this April in 11 year's and 6 months.
She played safely and won. She said "Computer played non-booked
move in middlegame, and blew up."
--------------------------
Hiroshi Yamashita
----- Original Message ----- From: "David Fotland" <
[email protected]>
To: <[email protected]>
Sent: Saturday, October 02, 2010 1:01 PM
Subject: [Computer-go] Exhibition match
Erica is playing a 12 yr old 1 dan professional, granddaughter of
Fujisawa
Shuko, with 6 stone handicap.
Video is live now at
http://www.ustream.tv/channel/the-jaist-20th-an-ni-ver-sa-ry-events-go
--------------------------------------------------------------------------------
_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
--
=========================================================
Olivier Teytaud -- [email protected]
TAO, LRI, UMR 8623(CNRS - Universite Paris-Sud),
bat 490 Universite Paris-Sud F-91405 Orsay Cedex France http://0z.fr/EJm0g
(one of the 56.5 % of french who did not vote for Sarkozy in 2007)
_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
