Reproducible, Reusable, and Robust Reinforcement Learning - Joelle Pineau

Explore a comprehensive lecture on reproducibility, reusability, and robustness in reinforcement learning delivered by Joelle Pineau from Facebook/McGill University at the Institute for Advanced Study. Delve into the reproducibility crisis in science, policy gradient methods in reinforcement learning, and the challenges of fair algorithm comparisons. Examine the intricate interplay of hyperparameters, performance measurement techniques, and the role of infrastructure in reproducibility. Investigate the myths and facts surrounding generalization in reinforcement learning, and understand the complexities of applying RL to real-world scenarios. Learn about the ICLR Reproducibility Challenge and gain insights into creating more reliable and robust reinforcement learning systems.

Intro
Reproducibility refers to the ability of a researcher to duplicate the results of a prior study....
Reproducibility crisis in science (2016)
Reinforcement learning (RL)
Adaptive neurostimulation
RL via Policy gradient methods
Policy gradient papers
Policy gradient baseline algorithms
Robustness of policy gradient algorithms
Codebase comparison
An intricate interplay of hyperparameters!
Fair comparison is easy, right?
How should we measure performance of the learned policy?
From fair comparisons...
How about a reproducibility checklist?
The role of infrastructure on reproducibility
Myth or fact?
Generalization in RL
Natural world has incredible complexity!
Natural world = RL simulation
Real-world video = RL simulation
Step out into the real-world!
ICLR Reproducibility Challenge Second Edition, 2019