Breakthroughs and the Future of (Deep) Reinforcement Learning
summary
Q learning centers on a one-line Q-update that fills a table or a neural function approximator and learns by experience replay rather than explicit models, demonstrated with a Python OpenAI Gym taxi example that needs roughly 100,000 episodes to converge. Deep Q variants beat humans on ~20+ Atari games in 2013, TD‑Gammon reached expert level in 1995, and AlphaZero in 2017 learns from rules only and dominates chess go shogi after weeks of heavy compute. Recent work applies RL to molecule search and simulated robot assembly.
AI-written summary · about
chapters (12)
- 0:00BIG DATA CONFEREI
- 1:42THE LEARNING SYSTEM
- 3:37A: PSYCHOLOGY, CLASSICAL CONDITIONING
- 5:26C: BRINGING IT TOGETHER: A TIMELINE OF RL
- 9:07WHY & WHO: SUMMARY
- 10:04ADDING AN ENVIRONMENT
- 15:09STEPS IN Q-LEARNING
- 16:49FILLING THE Q-TABLE
- 21:38MILESTONES: WINNING GAMES
- 24:10TRICKS
- 28:40GYM UPDATE FROM 2013 TO 2019 & BEYOND
- 30:35KEY TAKE HOME CONCEPTS
related talks
49mML: Teaching machine to learn from its own mistakes (...)
28mDeep Reinforcement Learning: theory, intuition, code
26mAnkur Narang; Senior Vice President – Technology & Decision Sciences at Yatra at Cypher2017- Part1/3
49mReinforcement Learning: Demystifying the hype
01h 17mTraining intelligent game agents using deep reinforcement learning
30mReinforcement Learning - ChatGPT, Playing Games & More
01h 11mFoundations of RL
01h 03mIEEE CoG 2019 - Industry Day - David Silver (DeepMind)