March 22, 2008: The first officially sanctioned victory of a computer over a professional player took place at the Paris Go Tournament.
The match pitted Catalin Taranu, a professional 5th Dan, against the program MoGo (TAO), on a 9x9 board. Combining innovative Artificial Intelligence with parallel multi-processors and message passing technology, the underlying algorithm is usable for a wide range of problems in addition to playing the game of Go.
The 2008 Paris Go Tournament, hosted the inaugural IA-Go Challenge 2008 – an Artificial Intelligence Go competition, organized by the Récitsproque corporation. The competition consisted of three matches on a 9x9 board, all with the same professional player opposing MoGo.
Summary of the matches
- During the first match, Catalin Taranu won "easily", as he put
it. Due to a problem with the link to the Bull computers, MoGo was
disconnected from the cluster at the Laboratoire de Recherche en Informatique computer science research laboratory at Université Paris-Sud.
- The program won the second match against the Go master,
who explained that his loss was due to the machine "taking perfect
advantage" of his mistakes. (The connection to the Bull computers had
been reestablished, and continued to operate for the rest of the
tournament).
- Finally, Catalin Taranu won the third match, after a particularly hard-fought battle.
Afterward, the players wanted to continue competing outside the official tournament, and MoGo proved invincible during additional matches against highly-ranked players.
>> Learn more about the matches
In this game, where humans were once considered to have an enormous advantage, the computer has managed to gain ground, now reaching the level of a professional Go player on a 9x9 board. The game of Go, even more complex than chess, is based on a set of possible moves that mathematically exceed the number of atoms in the universe...
How and why?
MoGo is built on three key scientific and technical advances:
- First, the so-called "multi-armed bandit" algorithm
allows partial exploration of the mathematical space of possible moves.
This approach has revolutionized the world of planning in indeterminate
problem spaces.
- Second, the evaluation of board positions uses Monte Carlo algorithms, which simulate the behavior of an inexpert, stochastic player, but without any prior knowledge (bias).
- Finally, parallel processing provides the computing power necessary to obtain the results of Monte Carlo calculations with sufficient precision.
These advances have permitted the application of this algorithm in numerous other domains, such as resource management.
Who?
Numerous improvements – including the Monte Carlo and parallel processing aspects – were made during 2007 and 2008, due to work by Olivier Teytaud, Jean-Baptiste Hoock, Arpad Rimmel and Julien Pérez, with the help of Thomas Hérault for the optimization of the parallel code (Grid 5000), Vincent Néri and Jean-Francois Méhaut.
Other improvements were made by Jean-Yves Audibert, Vincent Danjean and his MOAIS team, David Silver of the University of Alberta, and the entire TAO team.
MoGo benefitted from the contributions of several renowned Go players, particularly Frédéric Donzet and Bernard Helmstetter.
Finally, Alain Facélina, Marc Jégou and Eric Caudal of Récitsproque made the IA-Go tournament possible, with the help of Clément Trung.
Thanks also go out to the computer-go mailing list, the KGS server and the ongoing Cgos tournament.
>> Previous step: January 2007 MoGo is top ranked in the ongoing Go tournament (...for computers!)