AI beats best Stratego player 15-1 on a budget
Original: With most information hidden, the game Stratego had stumped AI until now
Why This Matters
Beating Stratego with cheap hardware expands AI's reach in imperfect-information domains.
A team from CMU, MIT, NYU, and Stanford built Ataraxos, an AI that beat Stratego world-class player Pim Niemeijer 15-1 with 4 draws, trained on just 16 GPUs for a few thousand dollars.
Chess fell in 1997, Go in 2016, poker years before that. Stratego held out—until now. Researchers from Carnegie Mellon, MIT, NYU, and Stanford have built an AI called Ataraxos that beat Pim Niemeijer, widely considered the best Stratego player ever, 15 games to one with four draws. The kicker: it cost only a few thousand dollars to train on 16 GPUs.
Stratego's resistance to AI comes down to scale and deception. Each player has 40 pieces of unknown identity; piece identities reveal only on collision. "There's 40 pieces on the board that could be in any order," said MIT co-author Gabriele Farina—more than a decillion possible setups. Games can stretch to 2,000 moves versus chess's typical 40. And bluffing is core: move a weak piece like a marshal to intimidate, but do it too often and opponents stop fearing you.
Earlier efforts like DeepMind's DeepNash (2022) failed partly because hidden information causes self-play training to loop. Ataraxos tackled this with adaptive learning rates—large strategic changes early in training, fine adjustments later, across 163 million self-play games. The key innovation was a second neural network that estimates the probable identity of hidden enemy pieces, enabling limited look-ahead search before each move. DeepNash lacked this capability entirely.