Tuomas Sandholm is a professor at CMU and co-creator of Libratus, which is the first AI system to beat top human players at the game of Heads-Up No-Limit Texas Hold'em. He has published over 450 papers on game theory and machine learning, including a best paper in 2017 at NIPS / NeurIPS. His research and companies have had wide-reaching impact in the real world, especially because he and his group not only propose new ideas, but also build systems to prove these ideas work in the real world. Video version is available on YouTube. If you would like to get more information about this podcast go to https://lexfridman.com/ai or connect with @lexfridman on Twitter, LinkedIn, Facebook, or YouTube where you can watch the video versions of these conversations.
This podcast episode features an interview with a game theory and machine learning expert who has published over 450 papers on the subject, including research on applying these concepts to the game of poker. His work has had real-world impact through the development of systems to test and implement his theories.
The speaker explores the possibility of playing poker for money against the top players in Heads Up No Limit Texas Hold'em, which is different from the multiplayer version. Four of the top 10 players were invited to participate in an event to play against AI for testing purposes.
The action abstraction technology in the beginning finds the optimal combination of bet sizes which is provably convergent, but not very scalable. Meanwhile, information abstraction concerns the abstraction of what chance does in terms of abstraction in general games, since the top players are so good at hiding tells that it is not worth trying to find them in each other.
The paper "Depth-Limited Search for Imperfect Information Games" introduces sound depth-limited local heads to solve imperfect information games, where learning an evaluation for a state is not enough due to dependence on both players' beliefs.
The concept of information sets in game theory allows for uncertainty in decision making, specifically when one player is unaware of the moves made by another player in the game. These information sets represent a range of potential states within the game, allowing for more nuanced strategies to be developed.
This podcast explains how the concept of stochastic and extensive form games are utilized in machine learning to gain data about an opponent and tweak a strategy.
This podcast episode explores the concept of modeling human behavior using game theoretic approaches and discusses how to make games like poker beyond current AI methods.
The computational game theory community is experimenting with AI systems in poker and chess. While AI systems are now outperforming humans, humans still love the games.
This podcast discusses the challenges in achieving certain objectives in mechanism design, and how even experienced designers still rely on theoretical insights. The guests also touch on the potential for automated mechanism design tools.
Computational game theory can be used to improve military and business strategies based on game theoretic methods such as Nash equilibria. It is an open problem that can excite researchers to expand these theories into real-world applications such as stock market trading.
Game theory has a pivotal role to play in addressing the gap between AI theory and reality. Although AI can optimize objectives to the hilt, if it has the wrong objective, it can cause more harm than good compared to a human with some insight and partial optimization.
The development of strategic machine and strategy robot that integrates data, computing power, and deep learning is vital in solving bigger games with hidden player actions. Even more improvements are needed in the technology to completely integrate them into the system.