Ataraxos beats Stratego champion 15-1-4 for under $8,000

A team of academic researchers built an AI that beat the most decorated human Stratego player of all time by a 15-1-4 margin over 20 games, and it cost under $8,000 to train. The system, Ataraxos, is described in a paper in Nature by authors at Carnegie Mellon, MIT, NYU and Stanford, who report the first superhuman result in Stratego's history.

The compute gap is the striking part. DeepMind's prior Stratego system, DeepNash, cost an estimated $3 million to $4.5 million to train. The authors say Ataraxos consumed roughly 1/500th the compute, 1/30th the self-play games, and 1/100th the training examples. The RL training ran on 16 NVIDIA H100 GPUs for a week; the belief network trained on 4 H100s for four days.

"Ataraxos defeated the most decorated human Stratego player of all time by a large margin," the paper reports, calling it "to our knowledge, the first superhuman result in the game's history." The human opponent was Pim Niemeijer.

The system has three pieces: a policy-value network trained by self-play, a belief network that models hidden opponent information, and a test-time search procedure. A technique the authors call "dynamic damping" coordinates regularization strength with policy update magnitude (stronger regularization and aggressive updates early in training, weaker regularization and smaller updates later), which the paper credits with stabilizing learning under imperfect information.

The recipe generalized. The same techniques produced a superhuman AI for Barrage Stratego, which the paper says defeated three multi-time world champions, plus new state-of-the-art agents for Hanabi across two- to five-player variants and the Chinese card game dou dizhu, beating prior bots PerfectDou and DouZero. The authors describe the result as "a design pattern for reinforcement learning and search that is effective under large amounts of hidden information."