Pim Niemeijer, the world’s top Stratego competitor, lost to an AI named Ataraxos by a score of 15, with four matches ending in a draw. The machine needed just 16 GPUs and a few thousand dollars to train.
For a long time, machines had trouble with the game. Then Deep Blue beat Garry Kasparov at chess in 1997, and AlphaGo beat Lee Sedol at Go in 2016. Poker bots have beaten professionals for years. Stratego, however, held out.
The game at a glance
Stratego is an imperfect-information game, just like poker. Each player gets 40 pieces — ranks from marshal down to spy, plus bombs and a flag. The goal is to capture the opponent’s flag. Your opponent knows where your pieces are, but not what they are. Identities are revealed only when two pieces collide in battle — the weaker one is removed, and the winner’s identity shows.
“There’s something super distinctive about Stratego, which is that it is a massive amount of hidden information that unfolds over a very long time scale,” said Eugene Vinitsky, a researcher at NYU and co-author of the study.
The hidden information
In some forms of poker, the hidden information is tiny. In Texas Hold’em, “You only have two hidden cards,” said Gabriele Farina, an MIT computer scientist and another co-author. That leaves just 1,326 possible hands, few enough for a machine to weigh them all.
“In Stratego, there’s 40 pieces on the board that could be in any order,” Farina said. That’s more than a decillion possible setups. Then there’s the game’s length.
“In chess, usually the game lasts 40 moves, but in Stratego, a game can easily last 2,000 moves,” Farina said.
The bluffing problem
In Stratego, players sometimes pretend a weak piece is a marshal to frighten an opponent away from a position. If a player uses this trick too much, his threats stop carrying weight. But if he never pretends at all, the other side can read his hand with ease.
The challenge of finding that balance was what earlier AIs such as DeepMind’s DeepNash, introduced in 2022, could not overcome.
The win
A team of researchers from Carnegie Mellon, MIT, NYU, and Stanford created Ataraxos, and it was Ataraxos that finally cracked it.
The resources
The training cost was modest. Just 16 GPUs and a few thousand dollars got the job done.
What this means
What makes the victory stand out is what it resolves. The program defeats the world’s top human player, and it also settles an issue that had previously eluded other AI systems.
“There’s something super distinctive about Stratego, which is that it is a massive amount of hidden information that unfolds over a very long time scale.”
The verdict
The win marks a turning point for AI study, far beyond its significance to those who play Stratego. It demonstrates what can be achieved by joining a vast search space with a lengthy timeline and concealed information. Previous efforts fell short because they could not manage the bluffing without sacrificing strength somewhere else.
A team cracked a puzzle that had DeepMind’s DeepNash stuck, doing so with almost no resources to speak of — just 16 GPUs and a few thousand dollars.
A quick look at the numbers
| Game | Year | Best player defeated |
|---|---|---|
| Chess | 1997 | Garry Kasparov |
| Go | 2016 | Lee Sedol |
| Stratego | Now | Pim Niemeijer |
The progression illustrates the advance of artificial intelligence. Chess was overcome first, followed by Go, then poker, and now Stratego. Every game presented a distinct challenge, yet each ultimately succumbed to a machine that grasped its principles more thoroughly than any human ever could.
Source material: “With most information hidden, the game Stratego had stumped AI—until now,” Ars Technica.
Get the Notebook.
The day's best stories and every fresh verdict, in plain English, in your inbox by seven. One email a day, no more.