Overblog Suivre ce blog
Editer l'article Administration Créer mon blog
4 mai 2011 3 04 /05 /mai /2011 15:18

 

  http://www.lri.fr/~teytaud/nc2.jpg

http://www.irisa.fr/myriads/software/grid5000/generalview

SMP clusters form NUTN (left) and

Grid5000 (right)

 

 

http://wcci2010.nutn.edu.tw/images/pcc.jpg

http://wcci2010.nutn.edu.tw/images/jxzhou.jpg

Ping-Chiang Chou

(5 Dan pro Go player)

 

 

 

 

Below the summary of games between computers and humans
in SSCI 2011; SGFs are here: http://ssci2011.nutn.edu.tw/result.htm
and Go from random initial board report is http://www.lri.fr/~teytaud/randomgo.pdf (draft)
and blind-Go report is http://www.lri.fr/~teytaud/blindgo.pdf (draft)

Pachi and MoGoTW are running on clusters of SMPs; MoGoTW and been supported by the Grid5000 project (www.grid5000.fr). It is based on a work in the Tao team (lri, inria),

completed by a joint research project with the National University of Tainan (Taiwan).

Chun-Hsun Chou
(9 Dan pro Go player)
1) 9x9 Blind Go: 2 pros players (Ping-Chiang Chou 5P and Chun-Hsun Chou 9P)
   played four 9x9 games against MoGoTW.
    C.-H. Chou 9P lost 2 games (one as black one as white) and
    P.-C. Chou 5P won 1 game out of 2. The lost game would have been a
      win in standard 9x9 Go.MoGoTW was running on Grid5000 for P.-C. Chou
     and on NUTN-Cluster (3 nodes/40 cores/ 100Gb RAM at NUTN, Taiwan)
     for C.-H. Chou.http://ssci2011.nutn.edu.tw/go_jpg/20110413/20110413nutngo1-mogotw1_3-1.jpg
The game won by MoGoTW as black against Chun-Hsun Chou
(9P, winner of LG-Cup 2007). The pro mentionned that move 9 was a very
good move and he knew he had lost at that point in the game.
    Computer wins 3 out of 4 games. The opening as black was significantly modified, compared to
   previous games of MoGoTW.

2) 13x13

    MoGo lost with H2 against P.C. Chou and C.H. Chou.
    MoGo won with H2.5 (H3 but reverse komi 3.5) against P.C. Chou 5P.
    MoGo lost with H2.5 (H3 but reverse komi 3.5) against C.H. Chou 9P:
http://ssci2011.nutn.edu.tw/go_jpg/20110413/20110413nutngo2-mogovspro2_5-2.jpg
First ever win of a computer against a professional
player in 13x13 Go with handicap 2.5
3) 19x19 Rengo
    MoGoTW and Pachi both lost one H6 game against P.C. Chou 5P and C.H. Chou 9P (humans playing in Rengo).

4) 19x19
   a) H7
    MoGoTW lost H7 against C.H. Chou 9P.
    Pachi won H7 against C.H. Chou 9P. The pro said that Pachi played pro-level for
       killing a big group:
http://www.lri.fr/~teytaud/pachiH7.jpeg
The win by Pachi with H7 against a top professional player
(the first time since MoGo's win in 2009)
   b) H6
    Pachi lost against P.C. Chou 5P.
    MoGoTW lost against C.H. Chou 9P.

5) Random initial positions against B. Helmstetter (french 5D, former french championship and former 4th in world amateur championship (usually winning with H6 against MoGo).
http://www.lri.fr/~teytaud/rg.jpeg
 
MoGoTW won both as white and as black
against B. Helmstetter from this initial position.
http://strasbourg.jeudego.org/Photos%202003/Helmstetter03.jpg
Advantage of Go from random initial positions:
  no tedious fuseki-learning. Try with children, it's
   always a success :-)
    MoGoTW played games with initial stones randomly distributed
on the board.  The positions were randomly drawn, with rejection when MoGoTW vs MoGoTW lead  to more than 28 or less than 22 wins out of 50.  Pie rule in favor of human, but in only a few minutes - human chooses black unless he believes there is a clear advantage for white.
Bernard Helmstetter; 5Dan amateur Go player,
4th in world amateur championship 2004

    No more than 160 stones: B Helmstetter wins everything (somehow easily).
    240 stones: computer wins first game as black. Replay changing colors:
            humans wins as black
    180 stones: human chooses to be white. Computer wins as black.
                      Replay changing colors: computers wins as white.
    180 stones, new position: human chooses his side and wins.
      (i.e. 1/2 for computer with 240 stones and 2/3 for computer with 180 stones
        ==> humans can compete against top level amateurs from random boards
        with sufficiently many boards; see e.g. http://www.lri.fr/~teytaud/rand180.sgf)

Partager cet article

Repost 0
Published by teytaud - dans Recherche
commenter cet article

commentaires

lisa 04/04/2017 10:29

It was worth reading it

Marry 22/03/2017 13:09

Nice
Thank you for sharing this