Post by @sunsetjesus • Hey

Performance on training games is very good, even from highly suboptimal data. With near optimal data, this outperforms non-Q-learning methods (e.g., BC, d