AI System Ataraxos Crushes World's Best Stratego Player 15-1, Trained on Just 16 GPUs

Ataraxos Defeats World Champion in Stratego

A collaborative team from Carnegie Mellon University, MIT, New York University, and Stanford has built Ataraxos, a game-playing AI that decisively beat Pim Niemeijer—arguably the best Stratego player in history—winning 15 games to one with four draws2

2

. Niemeijer holds four world championships and spent more than 600 weeks as the world's top-ranked player2

2

. The AI system for strategic decision-making achieved superhuman performance in Stratego using just 16 GPUs and a few thousand dollars in training costs2

2

, a stark contrast to DeepMind's DeepNash, which required millions in computational resources yet still couldn't reliably beat top human players3

3

.

Why Stratego Remained Unconquered

Stratego represents one of the most challenging imperfect information games for AI. Each player arranges 40 military pieces—from marshals to spies, plus bombs and a flag—on their side of the board2

2

. Opponents know where pieces are positioned but not their identities until they collide in battle. This creates more than a decillion possible setups—exponentially greater than chess, which has 10 to the 66th power possible piece configurations3

3

. "In chess, usually the game lasts 40 moves, but in Stratego, a game can easily last 2,000 moves," explained Gabriele Farina, an MIT assistant professor in EECS and senior author of the study2

2

. Eugene Vinitsky, an NYU researcher and co-author, emphasized what makes this distinct: "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"2

2

.

Technical Architecture Behind Scalable Decision-Making

Ataraxos—named after the ancient Greek word meaning calm and unbothered—employs two interdependent self-play reinforcement learning processes realized by transformer-based networks1

1

. The first handles set-up selection, where players privately determine starting positions. The second manages move selection during gameplay. The set-up network uses a decoder-only transformer architecture trained with Monte Carlo returns, while the move network employs an encoder-only architecture with λ-based expected return estimators1

1

. Ataraxos trained through 163 million self-play games, making big strategy changes early in training and small, careful adjustments later to avoid the circular learning patterns that plague AI for hidden-information games2

2

.

Breakthrough in Search and Belief Modeling

The critical innovation separating Ataraxos from DeepNash lies in its test-time search capability. "This is one of the things that we did figure out how to do," Farina noted2

2

. Ataraxos incorporates a belief model—a neural network trained on self-play games to predict opponent piece identities based on movement patterns1

1

. Instead of enumerating every possible arrangement, the AI samples plausible configurations, simulates candidate moves in each scenario, and selects based on outcomes2

2

. This combination of reinforcement learning and search proves powerful for strategic decision-making in imperfect information settings1

1

.

Calm Under Pressure and Strategic Bluffing

Ataraxos demonstrates distinctly non-human strategic behavior. "For humans, it's very hard when you know a secret to make decisions ignoring the fact that you know that secret," Farina explained. "For machines, it's easy"2

2

. When Ataraxos estimates opponents have no reason to suspect a weakness, it leaves vulnerable positions alone—moves humans would only make while bluffing. The AI beats Stratego player opponents by methodically working back from deficits rather than attempting desperate gambles. "We would watch the bot 'bluff' its way back from like a two percent victory probability, very, very casually," Vinitsky observed2

2

. Bluffing creates a delicate balance—too much renders threats meaningless, too little becomes predictable, a challenge that stumped earlier systems2

2

.

Real-World Applications Beyond Gaming

The implications extend far beyond board games. "In the kind of imperfect information tasks you would face in reality, you often don't have the luxury of enumerating through all the possibilities. There are just too many," Farina stated3

3

. The system could adapt to business negotiations, cybersecurity threats, military strategy, and financial trading—scenarios where parties possess information others lack3

3

. The success demonstrates that large amounts of hidden information are no longer prohibitive for practical AI applications, putting strategic decision-making tools within reach for any domain with fast, accurate simulators1

1

. Ataraxos also performed strongly in other strategic games with different rules, proving its generalizability3

3

.