Max Schwarzer
TalkRL: The Reinforcement Learning Podcast
00:00
The Future of Deep Resonance Learning
DER was, you know, the performance was terrible by modern standards. I think it gets an IQM score of about 0.2, which is one-fifth of what BBF does. But for 2019 it was a real innovation. Then there were a bunch of papers that came out after DER when people realized that sample efficiency was an open playing field and pretty much everyone knew we could do much better than what DER had shown. So then that was sort of where the field stood in 2020-2021. We introduced the idea of just resetting the neural networks parameters every once in a while. And it turned out that doing this improved performance really, really significantly on a
Transcript
Play full episode
