Scalable decision-making for games of imperfect information – Nature (www.nature.com)

In a groundbreaking development for artificial intelligence in games of imperfect information, a new AI named Ataraxos has been introduced, achieving unprecedented superhuman performance in Stratego. This AI, using innovative techniques in self-play reinforcement learning and test-time search, decisively defeated the most decorated human Stratego player, Pim Niemeijer, with a remarkable score of 15 wins, 1 loss, and 4 draws over a 20-game series. Significantly, Ataraxos accomplished this victory while requiring substantially less computational resources than previous efforts, showcasing a major advancement in both efficiency and effectiveness in AI strategy games. The success of Ataraxos not only in Stratego but also in Barrage Stratego, Hanabi, and dou dizhu highlights a new design pattern in reinforcement learning applicable to games with large amounts of hidden information. By employing a coordinated policy-value network and belief network during training, Ataraxos can effectively model uncertainties inherent to these games, allowing for robust decision-making strategies. This achievement marks a pivotal moment for the AI and machine learning community, as it addresses a long-standing challenge in strategic decision-making and broadens the potential for AI applications in various real-world scenarios, from finance to military strategy.