Artificial intelligence (AI) is fast making inroads into a variety of real-world strategic imperfect-information interactions, notably recreational applications such as games. In recent decades, AI has powered programs to defeat human experts in games such as checkers, chess and Go. However, unlike Poker, none of these games deal with hidden information, notably about opponents’ cards and their bluffing strategies. For the first time, researchers at Carnegie Mellon University developed an artificial intelligence entity known as Libratus that overpowered four top poker players at no-limit Texas Hold’em guided by head’s up rules.
The licensing of the AI technology developed has been earned by Strategic Machine, Inc., a company founded by Sandholm. The working of the AI is detailed in an article published online on December, 17 2017 in the journal Science.
AI Used Three-Pronged Approach and Supercomputing Ability to Simplify Poker
The 20-day two-player game competition was held at Rivers Casino in Pittsburgh, PA during January 2017. Developed by computer scientists Tuomas Sandholm and Noam Brown, Libratus used a modular approach and three-pronged strategy and computational methods based on Bridges computer at the Pittsburgh Supercomputing Center (PSC). The methods supported machine learning to identify lapses in opponent’s strategy while filling the loopholes in their own. The AI could garner points over $1.8 million in chips, which is equivalent to defeating the human experts by 147 mmb/hand.
Libratus to Prove Useful in Diverse Real-World Strategic Imperfect-Information Interactions
The AI break the entire game into computationally manageable parts, which enables them to exploit the opponents’ weakness and improvises in its strategy as the game proceeds to the final rounds. Notably, the researchers stated their AI used a blueprint strategy for the initial rounds of the game while for the final rounds this strategy guides finer-grained abstractions. In addition, the AI uses a strategy what developers call nested subgame solving.
According to researchers, Libratus used techniques that are not domain-specific. Hence, future can have potential applications in areas marked by hidden information such as business negotiations, cybersecurity, strategic pricing, and several military applications.