Ataraxos defeats top human Stratego player 15-1 in Nature

Ataraxos, a system from Carnegie Mellon, NYU, Stanford and MIT, beat the top-ranked human Stratego player 15 games to 1 with 4 draws, according to a paper published in Nature on 30 September.

The authors Samuel Sokota, Eugene Vinitsky, Hengyuan Hu, Zhiyuan Fan, J. Zico Kolter and Gabriele Farina put the win against Pim Niemeijer at an "85% effective win rate" and report that training the system cost "a few thousand dollars," compared with "$3,000,000 to $4,500,000" spent on prior Stratego efforts. The run consumed 163 million finished games and 208 billion environment steps.

Beyond Stratego, Ataraxos beat three world champions across four 50-game series of Barrage Stratego, scored 24.410 ± 0.009 on average with 58.05% ± 0.49% perfect games in five-player Hanabi, and outperformed PerfectDou and DouZero on the Chinese card game Dou Dizhu with statistical significance.

The architecture pairs a policy-value network trained by self-play with a belief network that infers hidden information, and refines its play at test time with a "40-ply 1,000-rollout search." The authors conclude that "reinforcement learning and search are no longer precluded from high performance by the presence of large amounts of hidden information."