Training RL From YouTube Videos
Reinforcement learning is great, but environment interaction can be expensive. This paper proposes an RL algorithm based of successor features that takes advantage of passive data to learn about the world without acting itself.
Outline
0:00 - Intro
1:41 - Offline-RL
2:50 - Successor Features
13:34 - Algorithm
21:11 - Results
27:50 - Criticisms & Thoughts
Social Media
YouTube - / edanmeyer
Twitter - / ejmejm1
Sources:
Paper - https://arxiv.org/pdf/2304.04782.pdf