Researchers develop AI that finally defeats world's best Stratego player
A team from Carnegie Mellon, MIT, NYU, and Stanford created Ataraxos, an AI system that defeated Pim Niemeijer, considered the greatest Stratego player ever, winning 15 games out of 20 with just 16 GPUs and a modest training budget. The breakthrough involved using a second neural network to guess the identity of hidden opponent pieces, solving the challenge of imperfect information that had eluded even well-funded AI labs. Stratego presented a more complex problem than chess or Go due to its massive hidden information, 2,000-move game length, and bluffing strategies.
Stratego has long resisted computational mastery because it combines several exceptionally difficult challenges. The game features an enormous decision space—over a decillion possible piece configurations—alongside gameplay that stretches across thousands of moves. Unlike poker variants where hidden information remains relatively contained, Stratego's concealed elements persist throughout extended play, requiring systems to maintain uncertainty management across extended sequences.
Ataraxos achieved its breakthrough by implementing a dual neural network architecture. While the primary network determined moves through self-play training across 163 million games, a secondary belief model continuously estimated opponent piece locations based on movement patterns. This allowed the system to sample probable scenarios rather than exhaustively evaluate possibilities, making forward-looking analysis computationally feasible in ways that had previously eluded better-funded competitors.
The achievement signals potential advances in AI systems handling imperfect information across domains beyond games. Researchers may apply similar belief-modeling techniques to real-world problems involving incomplete data, such as strategic planning or adversarial decision-making scenarios. However, the practical applications remain speculative; translating game-specific architectures to complex institutional or scientific problems could present substantial engineering challenges. The work primarily demonstrates algorithmic progress within computer science rather than immediate societal disruption.