[11] Jacob Andreas - Learning from Language
The Thesis Review
00:00
How to Beat Monazoom Is Revenge From Scratch
For the plan representations that were the sketch representations that we were using in the policy sketches paper there wasn't even that much structure. Really just I have a sequence of sub tasks the sub tasks have names the names tell you which tasks share sub tasks. And so the point there is just that this is an extremely simple kind of annotation to gather in addition to whatever sort of work you were building doing to build your reinforcement learning agent in the first place. So if I can you know spend. Three weeks tuning hyper parameters and it takes me five minutes to write down a text file that contains a little bit of information extra information about the structure of my problems. What can I actually do with the kind
Transcript
Play full episode
