Explore offline reinforcement learning and model-based optimization in this 34-minute lecture by Sergey Levine from UC Berkeley. Delve into the power of predictive models and automated decision-making, focusing on data-driven reinforcement learning and model-based optimization. Learn about off-policy RL, distribution shift challenges, and Q-function lower bounds. Examine the CQL algorithm and its performance. Investigate predictive modeling and design, addressing issues with simple prediction and exploring model-based optimization problems. Discover uncertainty and extrapolation concepts, and understand model inversion networks (MINS). Analyze experimental results and gain valuable insights into these cutting-edge machine learning techniques.
Offline Reinforcement Learning and Model-Based Optimization
Overview
Syllabus
Intro
What makes modern machine learning word
Predictive models are very powerful!
Automated decision making is very powerf
First setting: data-driven reinforcement lear
Second setting: data-driven model-based optimization
Off-policy RL: a quick primer
What's the problem?
Distribution shift in a nutshell
How do prior methods address this?
Learning with Q-function lower bounds Algorithm
Does the bound hold in practice?
How does CQL compare?
Predictive modeling and design
What's wrong with just doing prediction?
The model-based optimization problem
Uncertainty and extrapolation
What can we do?
Model inversion networks (MINS)
Putting it all together
Experimental results
Some takeaways
Some concluding remarks
Taught by
Simons Institute
