TalkRL: The Reinforcement Learning Podcast

Natasha Jaques

11 snips
Aug 9, 2019
Ask episode
AI Snips
Chapters
Transcript
Episode notes
INSIGHT

Deep RL Generalization Challenges

  • Generalization and robustness are major challenges in deep RL.
  • Small input shifts can cause policies to fail, limiting real-world deployment.
ANECDOTE

Inspiration from PhD Exam Pressure

  • Natasha Jaques struggled with a hard question from her PhD committee about social and emotional intrinsic motivations for agents.
  • She spent nearly 24 hours crafting the idea of rewarding agents for having causal influence on others' actions.
ADVICE

Dig Into Model Behaviors

  • Deeply analyze and visualize model behaviors beyond charts.
  • Audit RL policies carefully to uncover unexpected emergent strategies.
Get the Snipd Podcast app to discover more snips from this episode
Get the app