Ataraxos AI: Mastering Stratego with Belief Models and Efficient Training

Ataraxos AI achieves superhuman performance in Stratego

Researchers from Carnegie Mellon, MIT, New York University, and Stanford University have developed an AI called Ataraxos that has successfully beaten the best human Stratego player in history. In a series of 20 matches against four-time world champion Pim Niemeijer, Ataraxos won 15 games, lost only one, and drew four. The AI's dominance extended to the 2025 Stratego World Championship, where it won 38 out of 40 games against various challengers.

Stratego is an imperfect-information game where players must capture the opponent's flag while the identities of the opponent's 40 pieces remain hidden until they collide in battle. This creates a massive state space—exceeding a decillion possible setups—and games that can last up to 2,000 moves, making it significantly more complex for AI than games like poker or chess.

Solving the hidden information problem with belief models

The primary challenge in Stratego is the vast amount of hidden information that unfolds over a long time scale. Previous attempts, such as DeepMind's DeepNash (2022), struggled with the balancing act of bluffing and predictability. Ataraxos overcame this by introducing a belief model—a second neural network specifically trained to guess the identity of the opponent's hidden pieces based on their movement patterns.

Instead of attempting to iterate through every possible arrangement of pieces, the belief model allows Ataraxos to:

  1. Sample plausible arrangements of the opponent's army.
  2. Play out candidate moves within those sampled scenarios.
  3. Select the move with the highest probability of success.

This architecture enables the AI to maintain a state of "calm" (hence the name Ataraxos, from the Greek for unbothered), allowing it to make methodical recoveries from deficits that would cause human players to gamble impulsively.

Extreme compute efficiency and training speed

Ataraxos was trained with a fraction of the resources used by previous state-of-the-art models. While DeepNash required 1,024 specialized chips for two to three months (estimated at $3 million to $4.5 million in 2025 prices), Ataraxos was trained using only 16 GPUs for one week, with an additional four GPUs for four days to train the belief model.

This efficiency was achieved through two primary innovations:

  • GPU-accelerated simulation: The team wrote a simulator capable of running millions of moves per second directly on graphics cards, eliminating the bottleneck of CPU-based game logic.
  • Optimized self-play: Ataraxos played approximately 34 times fewer games than DeepNash (163 million total) while achieving superior strength. The team used a strategy of making large strategic adjustments early in training and smaller, refined changes later to avoid the "circular" learning patterns common in hidden-information games.

Impact on the Stratego metagame

The AI's approach to the game has already begun to influence human play. Ataraxos frequently employs a flag placement strategy—tucking the flag into a corner behind only two bombs—that was previously rarely used by human players.

Beyond Stratego, the researchers have demonstrated that this architecture is versatile. The same approach has been used to master the cooperative card game Hanabi, the Chinese card game dou dizhu, and Barrage Stratego (a faster eight-piece variant).

Future directions and real-world applications

The researchers argue that the techniques used to solve Stratego are applicable to real-world scenarios involving imperfect information and strategic negotiation, such as financial markets or military war-gaming. However, a current limitation is interpretability; Ataraxos cannot yet explain the reasoning behind its specific moves. The team's current research focus is on creating strategies that are both strong and explainable.

"For humans, it's very hard when you know a secret to make decisions ignoring the fact that you know that secret," says Gabriele Farina, an MIT computer scientist and co-author. "For machines, it's easy."

Sources

Related

  • Project
    lightvector/KataGo

    KataGo is a superhuman open-source Go bot trained via self-play that provides advanced game analysis, score estimation, and support for multiple hardware backends.

  • Dispatch
    Google DeepMind: 10 Years of AlphaGo's Impact

    Google DeepMind reflects on how AlphaGo's 2016 victory over a world champion catalyzed a decade of AI breakthroughs in science, mathematics, and the pursuit of Artificial General Intelligence (AGI).

  • Dispatch
    OpenAI Five

    OpenAI Five, a team of five neural networks trained via self-play on Dota 2, began defeating amateur human teams and showed that scaled-up reinforcement learning can handle long-horizon, partially observed tasks without fundamental algorithmic advances.

  • Dispatch
    NuminaMath 7B TIR wins AIMO Progress Prize – technical recap

    NuminaMath 7B TIR won the first AIMO Progress Prize by solving 29 of 50 hidden math problems, showcasing the power of a two-stage fine‑tuning recipe, large high‑quality math data, and a self‑consistency with tool‑integrated reasoning inference strategy.

  • Dispatch
    OpenAI Five: Dota 2 Reinforcement Learning

    OpenAI Five is the first AI system to defeat world champions in an esports game, demonstrating that large-scale self-play reinforcement learning can achieve superhuman performance in complex, continuous state-action spaces.