Chess had Deep Blue. Go had AlphaGo. Now, teams of human engineers have created a model called Ataraxos, which can dominate over human players at the game of Stratego.

If you’re like me, this conjures up images of playing with the little plastic pieces, the generals and the dukes and the bombs and the guys in tall hats, in an attic or basement, at your friend’s house, in the hazy days of the 1980s. (I forget the actual pieces, but there were bombs. They looked cool.)

Human children, many of those with the most access to the game, are more likely to muck about with the pieces than actually play. But Ataraxos is serious, and this virtual player, designed at MIT, Carnegie Melon, NYU and Stanford (it was a collaboration) can win. It beat the top human player in 2025, and a research paper released yesterday, September 30, shows how formidable this model is at Stratego, and why. It also explains why Stratego is its own form of complex game, compared to the two above examples, chess and Go. One fundamental difference is that the other games have a fully transparent game state, and Stratego, by virtue of those little plastic pieces with their faces turned inward, toward the owner, does not.

“Real-world decision-making generally involves hidden information, that is, information that is unknown to one agent but possessed by another,” the team writes in a paper published at Nature. “Unfortunately, the presence of large amounts of hidden information renders established reinforcement learning and search approaches ineffective. Even with multimillion-dollar industrial research efforts 1 , top-human-level play at Stratego—a board wargame with hidden information on a massive scale—has remained beyond the reach of artificial intelligence (AI).”

To beat this challenge, the engineers used something called “self-play reinforcement learning,” where, essentially, the machine examines a large number of games to extrapolate everything, including the bluffing and strategic elements.

That’s important, because while AI has been proving its supremacy in logic and calculated reasoning for years, it’s further behind on the intuitive stuff that we still feel is the province of human thought. But the new Stratego player is crafty.

“Ataraxos is good at calculating risk in a way that humans are not,” said one of the researchers, Gabriele Farina, an assistant professor in the Department of Electrical Engineering and Computer Science (EECS), as quoted in a contemporary MIT News piece . “A human might start freaking out if their most valuable piece is exposed, but the bot can be surprisingly composed. It doesn’t overcorrect and give away its secrets.”

This also explains the strange name. The name Ataraxos comes from the ancient Greek concept ataraxia (ἀταραξία), meaning roughly “untroubledness,” “serenity,” or freedom from disturbance. So, the game has a poker face.

It turns out that Ataraxos owes much of its design to the brilliant but chaotic mathematician John Nash, running on an algorithm called DeepNash that uses Nashian dynamics to compete with humans.

Technically, Ataraxos is a successor to DeepNash. It’s a different model. But in both cases, the machine learns through playing by itself. And it learns well.

“The success of this approach across adversarial, cooperative and team games establishes a design pattern for reinforcement learning and search that is effective under large amounts of hidden information, a longstanding desideratum of the field of strategic decision-making,” the paper authors write.

Interestingly, scientists have also designed AI for other complex games. One is Hanabi, a game developed in France, with a Japanese name. It’s a cooperative card game where players see everyone’s cards except their own. They must give limited clues, infer hidden information, and coordinate to build correctly ordered fireworks, making it a challenging AI benchmark.

Then there’s Dou Dizhu, a relic of the Chinese cultural revolution, where one “landlord” competes against two cooperating “peasants.” Players race to discard their cards using combinations, creating strategic challenges involving hidden information, cooperation, and competition.

Making AI players for games like these shows us some of that jagged frontier that Ethan Mollick and others talk about: AI getting scary-good at some things, but not others.

What about poker? That’s an American game where you bluff a lot, mixing the numerical realities of the 52-card deck (especially in the Omaha version of Texas Hold’em) with good old human trickiness. AI can do both now, and it turns out the reason that you haven’t heard of a single AI poker winner is that there are a lot of them out there. There’s Libratus, developed at Carnegie Mellon, which defeated elite human professionals in heads-up no-limit Texas Hold’em in 2017, and then there’s Pluribus, also developed by Carnegie Mellon, which lorded it over human players in six-player no-limit Hold’em in 2019. And it’s only got better since then.

The bottom line is that if you were planning on inviting AI to game night, regardless of what’s in your game cabinet, get ready to lose. Just kidding – to an extent. But the skill of these new models rests on some pretty strong foundations.