Thomas Sandholm: AI Game Theory & Solving Poker
The Breakthrough of Libratus
Professor Thomas Sandholm discusses the development of Libratus, the pioneering AI system designed to solve large-scale imperfect information games. Unlike other AI milestones like Chess or Go, poker introduces the critical challenge of hidden information, where players must act without knowing the full state of the game.
The Impact of AI on Poker
• Libratus defeated world-class human professionals in Heads Up No Limit Texas Hold'em.
• The research team overcame significant statistical variance by playing over 120,000 hands.
• The AI utilized automated, potential-aware abstraction rather than relying on deep learning, though modern iterations now incorporate learning methods.
Game Theory and Real-World Applications
Sandholm emphasizes that Game Theory is not just for tabletop games; it serves as a robust framework for complex real-world decision-making.
"I really believe in doing things in the real world and at scale. That's really where the proof is in the pudding."
Strategic Domains
• Autonomous Vehicles: Using automated negotiation to manage traffic merging more efficiently.
• Sourcing Auctions: Increasing supply chain efficiency through combinatorial mathematics.
• Security and Defense: Utilizing Strategy Robot to address cybersecurity and military planning challenges.
The Future of AI and Safety
While some experts voice concerns regarding existential threats and value misalignment, Sandholm remains optimistic. He argues that most of these risks are theoretical, whereas his real-world deployments—such as kidney exchanges—have proven to be demonstrably beneficial and safe. He advocates for a pragmatic approach: testing theories at massive scale to bridge the gap between mathematical models and practical utility.