At the Easter weekend, March 22nd/23rd 2008, as part of the Paris Go Tournament, the "IA-GO Challenge" was held. This was a series of games between Catalin Taranu and MoGo.
Catalin Taranu is a Romanian 5-dan professional. He is the strongest European-born Go player.
MoGo is a Go-playing program, using UCT. It was originally written by Sylvain Gelly and by Wang, Yizao, with ideas from Remi Munos, minor contributions from Olivier Teytaud, and help from Rémi Coulom; and subsequently improved by Arpad Rimmel, Jean-Baptiste Hoock, Julien Perez and Olivier Teytaud. Thomas Herault and Vincent Danjean have helped with profiling various aspects of the parallelisation. See a full list of contributors to MoGo.
It was meant to be running on a cluster of 32 eight-core 3GHz systems made available by Bull, but this only operated fully for the second and third games. For the first and fourth games, the powerful Bull cluster stopped working part way through the game and MoGo then ran on an ordinary 4-core system provided by LRI, Université Paris-Sud.
The first three games, played on the Saturday, were 9×9 go, 7½ komi, 30 minutes each sudden death, with colours for the first game decided by nigiri and alternating thereafter. The fourth game, played on the Sunday, was 19×19 go, 45 minutes each sudden death, with MoGo receiving nine handicap stones.
The games were played on a real board with a real clock. A person copied MoGo's moves from a computer screen to the board and pressed the clock for it, and Taranu's moves from the board to the screen.
The games were relayed on KGS, where the game records can be found in the accounts of both IaGoChall and teytaud. This site presents "cleaned-up" game records, incorporating the more useful kibitzes from both sets of game records and the less useful ones from neither (in my personal opinion). The four sub-headers below are links to these SGF-format game records. If you click on these you may, depending on your system settings, be able to download them or to play through them in your favourite SGF-reader. The times remaining to the players, as shown on these game records, are approximately right, maybe slightly too favourable to MoGo because of delays in pressing its physical clock.
MoGo played black in this game. Its 256-core cluster stopped working during move 13, and it then ran on the 4-core system. After this it never had much chance. Once Taranu played his move 16, establishing two secure groups, the game was hopeless for MoGo.
Taranu played black in this game. It seems (to this humble 2-kyu player) that he tried to establish three groups, and that MoGo's move 16 made it unlikely that all three were going to live. Nor could Taranu find an effective way to sacrifice one of them for the benefit of the others, and he eventually resigned.
MoGo played black again, and began as in game 1 – of course, it had no memory of its earlier loss. It diverged with its move 7.
After MoGo complicated the game with its move 15 (rather than go into a lost endgame) it looked as if either player could win. Taranu spent half of his remaining 22 minutes on his move 18, and it seems his analysis was accurate; MoGo could find no way to win after this.
This was a nine-stone handicap game. Again the 256-core cluster stopped working early in the game and MoGo was transferred to the 4-core system. The cluster recovered later in the game (at which move is not recorded) and MoGo was transferred back onto it.
It was interesting to see how a really strong player such as Taranu manages a nine-stone game against a player that does not make blunders. MoGo killed a group on the right side of the board early on, and at move 200, in the endgame, it seemed that it was still ahead; and it is good at the endgame. However, with move 208 Taranu showed his willingness to fight a ko (something MoGo does not excel at), and by move 215 it found itself fighting another, bigger ko. Taranu's move 242 effectively won this ko, bringing his dead group back to life, after which MoGo had no chance.
Overall, I think MoGo did well. No-one had expected it to win the 19×19 game; but even running on only a single computer, it had good chances well into the endgame. And of the two 9×9 even games for which its cluster was working, it won one out of two.
The following names appear in the game records:
This page is part of computer-go.info.