DeepNash AI Cracks Stratego, a Game of Hidden Information That Defeated Machines for Decades

DeepMind has built an AI system that plays Stratego at a level competitive with the game's best human players, marking the first time a machine has reached expert play in a game where most of the board is hidden from both sides. The system, called DeepNash, is described in a paper published in Nature and posted to arXiv under the title 'Mastering Stratego, the classic game of imperfect information.' DeepMind says the agent reached the top tier on the Gravon online Stratego server, earning a ranking comparable to the site's best human players, and won more than 97 percent of its games against existing Stratego bots.

Stratego is played on a 10x10 board with 40 pieces per side. Each piece carries a rank, and a player sees only the location of opposing pieces, not their identities, until a challenge forces a reveal. That means a single move can hinge on guesswork about what an unseen piece might be, and bluffs, feints, and long-term deception matter as much as raw calculation. Chess and Go, by contrast, give both players complete information, which is why search-based engines like AlphaZero could dominate them.

DeepNash does not rely on search over the game tree, the approach that powered earlier superhuman systems. Instead, it learns a strategy through self-play using a model-free reinforcement learning method, adjusting its policy against copies of itself across millions of games. That sidesteps a core problem: in imperfect-information games, the branching factor explodes once you account for what the opponent might be holding, and earlier attempts to bolt search onto Stratego tended to stall. The game has roughly 10 to the 535th power possible states, far larger than chess and comparable in some measures to Go, but with the added twist that players cannot evaluate a position directly.

The result matters beyond board games. Hidden information is the norm in real-world settings — negotiations, military planning, financial markets, and poker all turn on what an opponent knows and conceals. Poker AI such as Libratus and Pluribus already showed that machines can handle incomplete information in card games, but those systems leaned on explicit opponent modeling and game-specific abstractions. DeepNash's approach suggests a more general route: learn from scratch, without a hand-coded strategy library or a solver tuned to one game.

The paper's authors frame the work as a step toward systems that reason under uncertainty rather than perfect knowledge. That gap is where most practical AI still struggles, from autonomous driving to supply-chain planning. Whether DeepNash's self-play method transfers to messy real-world problems, or only to games with clean rules and bound state spaces, is the question the field will now try to answer.

The code and training details are available through the arXiv preprint, and the Nature paper includes an analysis of the agent's play style against human grandmasters. What remains unclear is how much of DeepNash's strength comes from the sheer volume of self-play games versus the specific architecture the team chose. Follow-up work will likely test whether the same recipe holds on other hidden-information games with different structures.