Reinforcement Learning - ChatGPT, Playing Games & More
summary
Reinforcement learning explains how agents learn from sequential interaction to maximize cumulative reward, with landmark results on Atari and AlphaGo and training that often requires millions to billions of episodes. Practical details covered include using Ray RLlib to run distributed RL, a CartPole demo that converges in ~50 episodes, offline reinforcement learning from logs and simulators (eg America's Cup rudder control), and RL with human feedback for ChatGPT using a reward model and PPO to fine‑tune language models for better recommendations and ad serving.
AI-written summary · about
chapters (9)
- 0:00Intro
- 1:52Agenda
- 3:18What is reinforcement learning?
- 5:08Video: http://youtu.be/Lu56xVlZ40M
- 6:02What is RL? continued
- 10:18Ray RLlib
- 17:36RL & ChatGPT
- 24:10RL for recommendations
- 29:08Outro
related talks
26mCan Reinforcement Learning Lead to AGI?
44mReinforcement learning: A gentle introduction
31mDesigning game AI with reinforcement learning
26mChatGPT and Minecraft's diamonds playing agents: Reinforcement Learning in a nutshell with sheeprl
28mDeep Reinforcement Learning: theory, intuition, code
02mDesigning a Video Game AI
49mReinforcement Learning: Demystifying the hype
19mTraining Agentic Reasoners