Ataraxos beats Stratego's top human 15‑1 after a week on 16 GPUs
A Carnegie Mellon-led team says its Ataraxos system beat Pim Niemeijer, the game's most decorated player, in a 20-game match on under $8,000 of compute.
Researchers from Carnegie Mellon, NYU, Stanford and MIT say their AI, Ataraxos, beat Stratego champion Pim Niemeijer 15 wins to 1, with four draws, in an official 20-game match. The Decoder reports the result appeared in Nature. Ars Technica covered it on Friday. Niemeijer has won four world titles and spent more than 600 weeks ranked first, according to The Decoder.
Stratego hides the value of each piece from the opponent, which makes it hard for search-based game AI. The Decoder puts the number of possible starting setups above 10^33. DeepMind's DeepNash reached the top three of the online ranking in 2022, but did not play the game's best player.
The team reports that Ataraxos trained for a week on 16 Nvidia H100 GPUs, plus four more days on four GPUs, at a cost under $8,000. The Decoder contrasts that with an estimated $3 million to $4.5 million and two to three months on 1,024 TPU nodes for DeepNash. Those comparisons are the authors' and the outlet's figures.
The system learns by self-play with no human game data. The Decoder says a key step was regularisation, which pushes the AI to vary its play instead of settling into predictable patterns. The authors are quoted as saying strong play requires a lot of randomness, because predictable players get exploited.
The authors say the same approach also worked in Hanabi, Dou dizhu and a Barrage variant of Stratego, and suggest it could help with strategic decisions in finance or negotiation. That is a claim of possibility, not a demonstration. No outside team has yet reported reproducing the training run or the match.