GoKawiilAtaraxos AI defeats Stratego world champion 15-1-4 in match series
Researchers from Carnegie Mellon, MIT, NYU, and Stanford unveiled Ataraxos, an AI system that defeated four-time Stratego world champion Pim Niemeijer in a series, winning 15 games, losing one, and drawing four. The system combines a self-play-trained strategy network with a second neural network that estimates the likely positions of an opponent's hidden pieces, allowing it to search a narrowed set of plausible game states before each move rather than attempting to brute-force the full hidden-information space.
GoKawiil's interpretation of the reporting above, not reported fact.
Stratego has long been considered a harder test for AI than games like chess or Go because key information is hidden from both sides, and prior systems such as DeepMind's DeepNash could not perform pre-move search due to the sheer size of the possibilities involving unknown pieces. By adding a belief-estimation network, the researchers suggest they found a general technique for narrowing hidden-information search that could extend to other imperfect-information domains such as negotiation or security games. The team also reports training Ataraxos on far less compute than DeepNash required, which, if representative, could indicate that strong hidden-information AI is becoming cheaper to produce.
- Ataraxos beat world champion Pim Niemeijer 15-1 with 4 draws in a Stratego match series.
- A new belief-estimation neural network lets the AI sample plausible hidden piece arrangements instead of brute-forcing them.
- The team estimates Ataraxos needed only 16 GPUs for a week, versus DeepNash's reported $3-4.5 million training cost.
Stratego Original Board Game — If AI besting a world champion at Stratego has you curious, there's no better way to appreciate the game's hidden-information depth than playing it yourself. It's a classic battle of deduction and bluffing that's easy to learn but endlessly strategic, perfect for family game night.
See Stratego Original Board Game on Amazon → Affiliate link — we may earn a commission on purchases, at no extra cost to you. Product picked by AI based on this article; it is not a tested recommendation.Source: games.slashdot.org, 2026-10-02
Published there as: “AI Has Finally Learned To Play Stratego”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.