Ataraxos beats top human Stratego player, 1/500 the compute

Ataraxos, a general-purpose AI for games of imperfect information built by a team from Carnegie Mellon, MIT, Stanford and NYU, beat Pim Niemeijer, described as the most decorated Stratego player in history, with 15 wins, 1 loss and 4 draws across a 20-game series. The result, published in Nature on September 30, was reached at roughly 1/500th the computational cost of prior efforts (DeepNash).

The authors call the 85% effective win rate "without precedent at the highest level of human play". Training ran on 16 NVIDIA H100 GPUs for one week, with a separate belief network trained on 4 GPUs for four days; the paper puts the full training cost at approximately a few thousand dollars.

The same recipe transfers. Ataraxos defeated three top-ranked Barrage Stratego world champions across four 50-game series, the first documented superhuman performance in that variant, and beat PerfectDou at Dou Dizhu by a role-averaged margin of 0.199. On cooperative Hanabi, it posted perfect-game rates of 77.53% in two-player and 89.90% in three-player games.

Three components carry the work: a policy-value network trained via self-play, a belief network that predicts opponent piece types from available information, and a search procedure that samples possible game states at test time. The authors credit a dynamic regularization schedule, stronger early and weaker later, with damping the chaotic learning dynamics that imperfect-information self-play tends to produce.

"The success of the AIs presented herein shows that the presence of large amounts of hidden information is no longer prohibitive," the paper concludes.