The State of Techniques for Solving Large Imperfect-Information Games
summary
Claims concrete wins and limits: a lossless abstraction reduces Rhode Island Hold ’Em from 3 billion nodes to a solvable LP with 1.2 million rows and columns and yields exact Nash, while two‑player limit Hold ’Em is solved to about 0.001 big blind per hand. Practical tools are potential‑aware card abstraction with Earth Mover distance, CFR/CFR+ with new regret‑based pruning, and first‑order (EGT) advances; a Claudico-style no‑limit bot plays 80,000 hands vs top pros and loses ≈0.091 mBB/hand (not statistically decisive).
AI-written summary · about
chapters (15)
- 0:00Introduction to game theory
- 1:18Applications and benchmarks
- 3:46Game abstraction techniques
- 7:35Practical abstraction algorithms
- 11:03Supercomputing and solvers
- 12:46Theoretical bounds and limits
- 17:10Game solving algorithms
- 21:11Comparing solver approaches
- 23:30Gradient methods and sampling
- 25:41Simultaneous solving
- 27:42Opponent exploitation
- 33:09AI versus human players
- 35:49Competition analysis
- 39:40Future research directions
- 43:18Q&A and final thoughts
related talks
46mThe Building Blocks of Superhuman Poker AI
47mRebel: Combining Deep Reinforcement Learning and Search for Imperfect-Information Games
01h 06mLearning in Games II
11mOm! Nom! Nash!
46m(Improving) Muddled Information
49mFast Rates for No-Regret Learning in General Games
20mThe Complexity of Infinite-Horizon General-Sum Stochastic Games
40mExtending Generalization Theory Towards Addressing Modern Challenges in Machine Learning