The French Federation of Go homologates a victory of a computer against a professional player in 9x9 Go.

english version

Saclay et Lille


Linux


Un des clusters utilisé pour spécifier MoGo (GDX, Grid5000).


The IAGO Challenge.
Since 2006, the BBMCP(1) technology has changed the state of the art in planification. MoGo is a BBMCP program for the specific case of the Game of Go. During the "Tournoi de Paris" (2), the company "Recitsproque" (6) and the FFG (3) organized the IAGO challenge, meeting between a professional Go player and a computer on a 9x9 Go board. The 9x9 Go is a game in which the superiority of humans on computers is very large.

The MoGo project
The MoGo project combines several teams (4), including teams working on parallelization (5), and uses clusters provided by Bull.

The 3 games
The first game has been disturbed by a technical problem on Bull's cluster. A standard computer has played the game instead of the cluster, and Catalin Taranu, 5th Dan Pro, winner of the Shusaku Cup 2008, won easily. During the second game, the cluster was back and MoGo won the game. Catalin Taranu said that he made some mistakes, and MoGo perfectly took benefit of these errors. The final position is as follows (MoGo is white):

The third game was played without technical trouble and Catalin Taranu won against MoGo with its full power. All the games are available here (SGF format).

Some other very high level players tried to play against MoGo and MoGo has not been defeated during these games.

Exhibition in 19x19.
MoGo also played a game in 19x19 against the same professional player, but Catalin won in spite of 9 handicap stones. The cluster from Bull has some troubles several times during the game and the game was partially played by standard machines. Catalin said that MoGo was close to the Dan level and played some brilliant moves; however, MoGo lost mainly due to a big error in the last part of the game.

The team thanks the FFG and Réciproque (6) for this opportunity of showing BBMCP (1), an innovative AI technique which can be applied in several domains and in particular for resource management and other planification tasks.

MoGo thanks Jean-Yves Audibert, Eric Caudal, Bertrand Chardon, Rémi Coulom, Vincent Danjean et MOAIS, Frédéric Donzet, Alain Facélina, Sylvain Gelly, Jean-Baptiste Hoock, Bernard Helmstetter, Thomas Hérault, Marc Jégou, Jean-François Méhaut, Rémi Munos, Vincent Néri, Julien Perez, Arpad Rimmel, David Silver, all the TAO team, Olivier Teytaud, Clément Trung, Yizao Wang, la mailing-list computer-go, KGS, Cgos, Linux.

  • (1) BBMCP: Bandit-Based Monte-Carlo Planning.
  • (2) Tournoi de Paris: http://paris2008.jeudego.org/
  • (3) Fédération Française de Go: http://ffg.jeudego.org/
  • (4) Tao, Inria-Saclay, Cnrs, Lri, Université Paris-Sud, Cmap, Sequel.
  • (5) Grid5000, Equipe Parallélisme du Lri, MOAIS.
  • (6) Récitsproque