(2021), still very interesting. Especially the "post-overfitting" training strategy is unexpected.
The low sample efficiency of RL is well explained.
alt Hacker News
(2021), still very interesting. Especially the "post-overfitting" training strategy is unexpected.