A force of 40 pieces, each bearing a concealed rank from marshal down to spy, has at last been deciphered by a low-cost machine. The contest is Stratego, and the program that solved it is named Ataraxos. It defeated Pim Niemeijer, regarded as the greatest Stratego player of all time, 15 matches to one, with four draws.
Stratego has stood firm against the rise of machines for decades. Deep Blue’s triumph over Garry Kasparov at chess came first, back in 1997. Then AlphaGo bested Lee Sedol at Go in 2016. Poker bots have been beating professionals for years. Yet Stratego resisted, with even DeepMind, despite its vast budget, failing to build a machine that reliably bested the top human players.
A team from four universities — Carnegie Mellon, MIT, New York University, and Stanford — has now produced it. They trained Ataraxos using just 16 GPUs and a few thousand dollars.
What Stratego Hides
Stratego is a game built on lies. Every side starts with 40 pieces, each standing for a rank from marshal all the way down to spy, along with bombs and a flag. The goal is simple: take hold of the enemy’s flag. What makes it work is that your rival can see where your pieces sit, but has no idea what any of them actually is.
Two players do not know what their opponent holds until their pieces come together in combat. The piece with the lower rank is set aside, and the identity of the victor becomes known to all. This is why Stratego counts as an imperfect-information game, much like poker, which machines were able to solve long ago. “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,” noted this point in his remarks, Eugene Vinitsky, who is a researcher at NYU and co-author of the study.
The comparison to poker stops at the surface. 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 is more than a decillion possible setups. The game also lasts far longer than chess. “In chess, usually the game lasts 40 moves, but in Stratego, a game can easily last 2,000 moves,” Farina said.
The Bluffing Problem
Then there is 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. On top of that, Stratego is a game of bluffing. Sometimes you move a weak piece as if it were a marshal, just to scare the opponent off.
If a player bluffs too much, their threats lose all weight. If they never bluff at all, they become easy to read. It is that careful balance, the team says, that earlier AIs like DeepMind’s DeepNash, introduced in 2022, found hard to manage.
A huge amount of hidden information, an extreme length, and constant deception made Stratego a harder target than chess or Go. The game demands a balance of calculation and psychology that earlier systems could not manage.
Ataraxos Wins
The Ataraxos team built a solution that required far fewer resources than DeepMind’s earlier work did. They trained their AI on 16 GPUs and spent a few thousand dollars, yet managed to match the performance of DeepMind’s prior attempts.
Niemeijer is widely considered the best Stratego player of all time. His record speaks for itself. The team’s result was 15 wins to one, with four draws. That is a dominant showing against a player who has spent his career studying the game’s hidden depths.
Why It Matters
Stratego has finally fallen, and the win carries weight because it closes a gap in AI’s record. Earlier games gave way to machines: Deep Blue defeated chess, AlphaGo overcame Go, and poker bots long ago began beating human professionals. Stratego, however, proved more resistant, its immense hidden information and extraordinary length making it a harder target than any game before it.
This victory demonstrates what a small team can accomplish. Though the researchers hailed from four major universities, their resources remained modest: they trained Ataraxos on 16 GPUs and spent a few thousand dollars.
This outcome marks a turning point. A game that stood firm against AI for many years has at last given way, and it happened with almost no money behind it.
The Numbers Behind the Win
- Game: Stratego
- Winner: Ataraxos (AI)
- Loser: Pim Niemeijer (arguably the best Stratego player of all time)
- Score: 15 wins to one, with four draws
- Training hardware: 16 GPUs
- Training cost: A few thousand dollars
- Earlier AI attempt: DeepMind’s DeepNash, introduced in 2022
What We Make of It
AI progress does not move along a straight line. Some problems get solved fast, while other issues hold firm for years. The game Stratego was one of those difficult cases. It has now been resolved, and the answer came from university researchers instead of a well-funded company.
A small setup yielded a strong result, which is a lesson for anyone building AI systems: sometimes less is more. The team’s approach was efficient, producing a strong result with minimal resources.
The victory demonstrates that AI research is growing more open to all, with a modest sum of money and 16 GPUs now well within the grasp of numerous labs.
This victory marks a significant achievement, yet there remains much work to be done. Many games still conceal their secrets from us, and the methods created for Ataraxos could serve as a foundation for the research that follows, helping investigators approach these remaining puzzles.
The group has earned praise for resolving an issue that left DeepMind searching for answers. Their achievement shows that artificial intelligence remains a relatively new area, where unexpected results keep turning up.
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.