MIT News breed · Science

New AI Masters Stratego, Outperforming Humans

Researchers have developed an AI system that has achieved superhuman performance in the complex board game Stratego, a feat previously unattained by AI. This advancement could have implications for human decision-making in scenarios with hidden information.

The AI, named Ataraxos, was created by a collaboration of researchers from MIT, Carnegie Mellon University, New York University, and Stanford University. It defeated top-ranked human Stratego players by a significant margin, utilizing advances in machine learning tailored for games with hidden information.

Stratego, a two-player game where opponent piece identities are concealed, is a challenging benchmark for AI strategic thinking. The number of possible piece configurations in Stratego far exceeds that of chess, making it particularly difficult for AI.

Ataraxos combines efficient training algorithms with new techniques for decision-making under uncertainty. It achieved superior performance to previous AI models while being less computationally demanding and cheaper to train.

The AI system demonstrated its generalizability by also outperforming top human players in other strategic games with different rules, suggesting its potential for various real-world applications.

The researchers adapted Ataraxos for other imperfect information games like Barrage Stratego, Hanabi, and Dou dizhu, achieving superhuman performance in each case. This highlights the method's broad applicability.

Future work aims to incorporate interpretability measures into Ataraxos, enabling the system to explain its decision-making process to humans. This is crucial for human oversight and trust before the AI can be adopted for real-world problems such as business negotiations or cybersecurity.

AI-samenvatting op basis van de bron.