Ataraxos AI Beats Stratego World Champion 15-1 With Four Draws
Ataraxos adds an inference step that DeepNash lacked, searching the opponent's hidden setup during play rather than only selecting moves from a self-play network.
Reporting from 1 source: GIGAZINE.
A research team from Carnegie Mellon University, MIT, and other institutions built an AI called Ataraxos that beat four-time world champion Pim Nijmeijer at Stratego, winning 15 games, losing 1, and drawing 4. The team says this is superhuman performance in a game of imperfect information. Ataraxos learns through self-play and infers the opponent's hidden pieces during play.
Stratego hides piece identities from both players, so the AI must infer what it is facing while it plays. The team's approach pairs a policy-value network built through self-play with a belief network that reasons about the opponent's hidden pieces. DeepNash, by contrast, did not search future developments during play. The reported match against Nijmeijer ended with fifteen wins, one loss, and four draws.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.