Libratus is an artificial intelligence (AI) developed by Carnegie Mellon University's Computer Science Department, specifically by Professor Tuomas Sandholm and his Ph.D. student Noam Brown. It is notable for its achievements in playing poker, particularly No-Limit Texas Hold'em, a game of imperfect information.
Key Achievements: Libratus gained international recognition in January 2017 when it participated in and won a 20-day, 120,000-hand poker tournament against four of the world's top professional No-Limit Hold'em players: Dong Kim, Jason Les, Daniel McAulay, and Jimmy Chou. The tournament, "Brains Vs. AI: Libratus," was held at Rivers Casino in Pittsburgh. Libratus ended the competition with a significant lead, earning virtual chips equivalent to $1.76 million, demonstrating its ability to outperform human experts in a highly complex strategic game.
Technology and Development: Libratus employs advanced AI techniques, including a combination of game theory, search algorithms, and self-play reinforcement learning. Unlike previous poker AIs that often relied on pre-computed strategies for simplified versions of the game, Libratus was designed to compute strategies in real-time for the full complexity of No-Limit Hold'em. Its core approach involves:
- Counterfactual Regret Minimization (CFR): An iterative algorithm for finding approximate Nash equilibrium strategies in extensive-form games.
- Online Search: During gameplay, Libratus uses a sophisticated search algorithm to refine its strategy for the current subtree of the game, taking into account the specific actions observed.
- End-game Solving: For parts of the game with fewer remaining decisions, Libratus can compute extremely detailed, near-optimal strategies.
- High-Performance Computing: The AI utilized considerable computational resources, including supercomputers, to train its models and execute its complex decision-making processes.
Significance: Libratus's victory represented a significant milestone in artificial intelligence research, particularly in the domain of imperfect information games. Unlike perfect information games like chess or Go, where the entire state of the game is known to all players, poker requires reasoning about hidden information (opponents' cards), bluffing, and probability. Libratus's success indicated a major step forward in AI's ability to handle uncertainty, deception, and strategic thinking in real-world, dynamic environments, with potential applications beyond games in areas such as negotiation, cybersecurity, and medical diagnosis.