Scaling RL

This lab maybe?

• https://cmu-aire.github.io/pages/publications.html
  • Offline QLearning on Diverse MultiTask Data Both Scales And Generalizes

Read this blog https://www.interconnects.ai/p/scaling-rl-axes

“In generative modeling, cross-entropy loss improves smoothly with model size and training compute, following a power law plus constant scaling law…”

There’s also Seohong Park’s blog that addresses some pretty important points about how to scale offline RL, and how this is an open-ended problem:

Papers:

• Horizon Reduction Makes RL Scalable
• Value-Based Deep RL Scales Predictably