Stratego Falls to AI After Years of Stumping Game Bots

Stratego has resisted the kind of AI breakthroughs that conquered chess and Go. Researchers now report that a system pairing two neural networks has made real headway on the board game, thanks to a simple but powerful addition: a model whose only job is to guess what the opponent's hidden pieces are.

The challenge is built into the rules. In Stratego, each player commands an army of pieces whose ranks are concealed from the other side. You only learn what a piece is when it attacks or is attacked. Chess and Go are games of perfect information, where both players see the entire board and a program can search deep into the future. Stratego offers no such luxury, and the number of possible hidden-state combinations explodes far beyond what traditional search can handle.

That uncertainty is why the game has been a long-standing benchmark for AI researchers studying imperfect-information problems. Poker bots found ways around hidden cards, but Stratego is far larger and runs far longer, with games lasting hundreds of moves. Earlier systems tended to play passably but fell short of the best human players.

The new approach splits the problem in two. One neural network handles decision making, choosing moves. The second is dedicated to inference, estimating the probable identity of each unseen enemy piece based on how it has moved and behaved. Feeding those guesses back into the decision maker gives the AI something close to the intuition human players use when they decide that a piece advancing aggressively is probably high ranking, or that one sitting still might be a bomb.

In effect, the system learns to reason about what it cannot see rather than treating the board as an impenetrable fog. By narrowing the huge space of possibilities into a manageable set of likely scenarios, the planner can make sharper choices without needing to enumerate every hidden configuration.

The result matters beyond the game. Many real-world tasks involve incomplete information, from negotiation and cybersecurity to logistics and autonomous driving, where an agent must act while uncertain about what others know, want or are about to do. A method that explicitly models hidden state and uses those beliefs to guide action offers a template that could transfer to such settings.

Caveats apply. Beating a benchmark game does not mean the technique will drop neatly into messier environments, where the hidden variables are not limited to a few piece types on a 10 by 10 grid. Still, cracking a game that has frustrated researchers for years is a meaningful marker, and it reinforces a growing lesson in AI research: when you cannot see everything, teaching a model to guess well can matter as much as teaching it to plan.