A team from MIT, Carnegie Mellon, NYU and Stanford has beaten the best human Stratego players by focusing on planning at the moment of play rather than on vast amounts of self-learning.
Published in Nature on 30 September, Ataraxos recorded 39 wins for 2 defeats against the Stratego elite at the World Championship, and a score of 15 wins, 1 defeat and 4 draws against the player presented as the strongest on the planet. Stratego is a puzzle for AI: each side ignores the identity of the opposing pieces, which prohibits exhaustive research that has made chess or go programs successful.
Less calculation, more reflection at the right time
Where DeepNash, the DeepMind system, relied on massive self-learning, Ataraxos combines a basic strategy from the game against itself with a planning phase at the time of each blow. A generic model considers the plausible identities of the opponent's hidden pieces, which avoids considering all possible configurations. The result: a level higher than DeepNash with less than one hundredth of his training examples and less than one thirtieth of his auto-game games.
The authors — Samuel Sokota (Carnegie Mellon), Gabriele Farina (MIT), Eugene Vinitsky (NYU), Zico Kolter (Carnegie Mellon), Hengyuan Hu (Stanford) and Zhiyuan Fan (MIT) — also report superhuman performances to the Stratego Dam, Hanabi and the dou dizhu, from games with very different rules and logics of cooperation.
Read the full article on the original source
Go to the source →