Ankur Narang: Senior Vice President – Technology & Decision Sciences at Yatra at Cypher2017- Part2/3
summary
Covers how environments map to Markov decision processes and how value iteration and Q-learning update rules become intractable for high-dimensional pixel inputs, so practitioners replace huge Q-tables with a deep Q-network that takes 4 stacked 84×84 frames and outputs ~18 action Q-values; shows Atari results that match or exceed humans, explains asynchronous actor-critic methods that run A3C 10x faster, and demos physics control using MuJoCo and combined policy/value nets.
AI-written summary · about
chapters (11)
- 0:00Start
- 0:06Markov Decision Process
- 5:15Value Function
- 7:00Deep Reinforcement Learning
- 8:55Fundamental Algorithm
- 11:14What Is the Challenge in Playing Atari Game
- 11:54Cue Table
- 16:02The Deep View Network
- 16:27Results of Atari
- 22:09Physics Simulation
- 25:58Alphago
related talks
28mDeep Reinforcement Learning: theory, intuition, code
01h 03mIEEE CoG 2019 - Industry Day - David Silver (DeepMind)
33mBreakthroughs and the Future of (Deep) Reinforcement Learning
49mML: Teaching machine to learn from its own mistakes (...)
17mTraining My Rival in Java: A Deep Q-Learning AI to Play Azul
21mQuantum Reinforcement Learning for Aerospace
21mTraining My Rival in Java: A Deep Q-Learning AI to Play Azul
52mThe Journey of AI in Games: A Technical and Philosophical overview