ML: Teaching machine to learn from its own mistakes (...)
summary
Covers concrete reinforcement learning methods and outcomes: Q-learning with Bellman updates and its practical limits using a Q table, then neural networks as function approximators to produce a policy; a Deep Q-Network trained on raw pixels learns Atari Breakout, after two hours playing like an expert and after four hours discovering the tunnel trick. Describes evolutionary neural-network genomes where simple creatures reach competent locomotion after ~80 generations, and cites state of the art: AlphaStar beating pros and OpenAI Five trained on 100k CPUs learning coordinated 5‑player Dota.
AI-written summary · about
chapters (10)
- 0:00Introduction to reinforcement learning
- 1:41Core concepts of reinforcement learning
- 4:35Q-learning fundamentals
- 13:39Run Jump Duck game demo
- 20:45Neural networks for Q-learning
- 27:50Deep Q-learning in Atari Breakout
- 29:38Evolutionary algorithms overview
- 36:38AlphaStar Starcraft 2 demo
- 42:21OpenAI Five Dota 2 demo
- 47:39Resources and conclusion
related talks
33mBreakthroughs and the Future of (Deep) Reinforcement Learning
01h 03mIEEE CoG 2019 - Industry Day - David Silver (DeepMind)
44mReinforcement learning: A gentle introduction
56mAn Introduction to Reinforcement Learning - Snake Your Way Out of a Paper Bag
01h 31mReinforcement Learning- Pac-Man
01h 01mReinforcement Learning: Pac-Man
28mDeep Reinforcement Learning: theory, intuition, code
26mHow Machine Learning Can Help