Ataraxos, an AI built to play Stratego, won 15 games against champion Pim Niemeijer in a 20-game series, with one loss and four draws, researchers report in a Nature paper published Sept. 30.
Stratego makes the opponent’s piece identities a central uncertainty. Each player secretly arranges 40 pieces, concealing their identities from the other player. Battles reveal the pieces involved. Ataraxos combines learning through games against itself with a network that models this hidden information and search that evaluates possible moves, the paper explains.
Playing with hidden pieces
The researchers trained separate, connected networks for arranging pieces and choosing moves. A belief network predicts the types of the opponent’s concealed pieces. Before a move, the system samples possible board states and simulates candidate actions to estimate their value.
The match series lasted three weeks, giving Niemeijer time to prepare between games and look for weaknesses. He was told that Ataraxos would not adapt to his play, according to the paper.
The reported 85% effective win rate treats each draw as half a win. Ataraxos won 15 of the 20 games outright, so the effective metric should not be read as an 85% share of victories.
What the result establishes
The authors describe the performance as, to their knowledge, Stratego’s first superhuman result. That is their assessment of the evaluation, rather than a direct comparison against every previous system.
They sought a match against DeepMind’s DeepNash, but report that DeepMind declined because its code was no longer functional. The paper therefore offers no head-to-head result between the two AIs.
