ReBeL - Combining Deep Reinforcement Learning and Search for Imperfect-Information Games

Explore a comprehensive video lecture on ReBeL, a groundbreaking algorithm that combines deep reinforcement learning and search for imperfect-information games like poker. Delve into the intricacies of this AI breakthrough, which achieves what AlphaZero did for chess and Go. Learn about the challenges of applying self-play reinforcement learning and tree search techniques to games with incomplete information. Discover how ReBeL converges to a Nash Equilibrium and creates a superhuman Heads Up No-Limit Hold'em bot with minimal domain knowledge. Follow along as the lecture covers topics such as game theory, belief representations, and the ReBeL algorithm itself, complete with theoretical foundations and experimental results. Gain insights into the broader impact of this technology on AI research and its potential applications beyond gaming.

- Intro & Overview
- Rock, Paper, and Double Scissor
- AlphaZero Tree Search
- Notation Setup: Infostates & Nash Equilibria
- One Card Poker: Introducing Belief Representations
- Solving Games in Belief Representation
- The ReBeL Algorithm
- Theory & Experiment Results
- Broader Impact
- High-Level Summary