Data Skeptic

Goodhart's Law in Reinforcement Learning

Mar 5, 2021
Ask episode
AI Snips
Chapters
Transcript
Episode notes
INSIGHT

RL Often Lacks Explicit Causality

  • Reinforcement learning appears tied to experiments but often omits explicit causal modeling.
  • This omission raises questions about why RL succeeds without formal causality handling.
ADVICE

Model Causality, Don’t Rely On Deep Nets

  • Do not assume deep networks automatically solve causal reasoning by magic.
  • Treat causal structure explicitly rather than relying on implicit representation inside networks.
ADVICE

Use Simple Models To Reveal Failure Modes

  • Use toy models to reveal failure modes and inspect learned policies exhaustively.
  • Enumerate states when possible to interpret networks and find causal errors early.
Get the Snipd Podcast app to discover more snips from this episode
Get the app