AI program beats world's top Stratego player 15 games to one

Researchers from Carnegie Mellon, MIT, New York University and Stanford built an artificial intelligence system that beat one of the best Stratego players in history, according to a study published in Nature. The AI, called Ataraxos, defeated Pim Niemeijer 15 games to one, with four draws, in a result the team says is the first time a machine has reliably beaten a top human player at the game.

Stratego is a board game where each side places 40 pieces representing military ranks face down, so the ranks are hidden from the opponent until two pieces collide in battle. That makes it an imperfect-information game, meaning players must act without seeing all the relevant facts, like poker. But the hidden information in Stratego dwarfs poker: there are more than a decillion possible ways to arrange the pieces, and games can run to 2,000 moves, researchers said. Those factors defeated earlier attempts, including DeepMind’s DeepNash, introduced in 2022.

Ataraxos learned by playing 163 million games against itself, reinforcing moves that led to wins and discouraging moves that led to losses. According to the research team, the key addition over DeepNash was a second neural network, called a belief model, that estimates what pieces an opponent is likely holding based on how those pieces have moved. Instead of accounting for every possible arrangement before each move, Ataraxos samples plausible ones and tests candidate moves against them.

The team reports that training cost a few thousand dollars, using 16 GPUs for about a week plus four more GPUs for four days to train the belief model. By contrast, the researchers estimate that DeepNash’s training run, on 1,024 specialized chips over two to three months, would cost $3 million to $4.5 million at 2025 prices. Ataraxos also needed about 34 times fewer training games than DeepNash to reach a stronger level of play, according to the team.

Niemeijer, who the researchers describe as holding four world championships and more than 600 weeks as the top-ranked player, played 20 online games against Ataraxos over three weeks and won once. Researchers say that single loss reflects the role of luck in Stratego rather than a flaw in the system, since even an optimal strategy loses occasionally. At the 2025 Stratego World Championship, Ataraxos won 38 of 40 games against other challengers. Gabriele Farina, an MIT computer scientist and study co-author, said machines can act on hidden information without the hesitation a human feels when concealing a weakness. “We would watch the bot ‘bluff’ its way back from like a two percent victory probability, very, very casually,” co-author Eugene Vinitsky said.

The same approach also beat three world champions at Barrage Stratego, a faster eight-piece variant of the game, mastered the cooperative card game Hanabi, and beat the strongest bots at the Chinese card game dou dizhu, according to the researchers. The team says its next goal is making the system’s reasoning understandable rather than just strong. Farina said Ataraxos cannot yet explain why it makes the moves it makes, and the researchers are now working on versions that stay strong while producing strategies a person could interpret.