Overview
Explore offline reinforcement learning in this 59-minute lecture by Emma Brunskill from Stanford University, presented at the Theory of Reinforcement Learning Boot Camp hosted by the Simons Institute. Delve into theoretical questions, notation, and key concepts such as generalization, covariate shift, and value error. Examine common tasks, assumptions, and evaluation criteria in batch reinforcement learning. Engage with discussion questions on important sampling, unbiased policy estimates, and models. Gain insights into doubly robust estimators and critically analyze the potential for wishful thinking in this approach to machine learning.
Syllabus
Introduction
Reinforcement Learning
Theoretical Questions
Notation
Outline
What if
Generalisation
Covariate Shift
Value Error
Common Tasks
Common Assumptions
Evaluation Criteria
Discussion Question
Important Sampling
Is this a form of wishful thinking
Unbiased policy estimates
Models
doubly robust estimators
Taught by
Simons Institute