The Building Blocks of Superhuman Poker AI
summary
Explains how Pluribus reaches superhuman six‑player no‑limit Hold’em by combining counterfactual regret minimization to learn a compressed blueprint strategy via self‑play with lightweight real‑time search (nested subgame solving) to refine moves at each betting round. Claims: beats top pros decisively (~50 milli big blinds per game). Resource notes: training costs about $150 on AWS (≈12k CPU hours), online play runs on two CPUs with ~128 GB memory. Includes a Python/C++ prototype and interactive demos that validate the approach.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 1:05About Skymind
- 3:50Poker AI
- 9:20What makes poker so hard
- 16:06How to approach imperfect information
- 17:41Regret matching
- 25:19Building variables
- 33:10Demo
- 33:35Testing
- 36:30Debugging
- 37:00Confusion
- 38:15Question
related talks
44mThe State of Techniques for Solving Large Imperfect-Information Games
47mRebel: Combining Deep Reinforcement Learning and Search for Imperfect-Information Games
46mBeating Go thanks to the power of randomness
08mMinimax search and the structure of cognition!
31mDesigning game AI with reinforcement learning
17mTraining My Rival in Java: A Deep Q-Learning AI to Play Azul
21mTraining My Rival in Java: A Deep Q-Learning AI to Play Azul
01h 07m(How) Are We Staying Smarter Than AI?