@LauraRuis @rockt Great thread, thanks for sharing! My setup is similar and I agree regarding the need for having a good note-taking system. I personally use Joplin and keep all my papers in Mendeley (& read on iPad with Papership). I started doing this 3 months ago and it has helped me a lot! I don't have the same note structure as you, instead I like to make survey notes explaining various topics, and notes about anything I come across. This whole system works like a second memory for me.

@surprisal I am still a little confused by the proof, although I know from GAE paper that both forms are acceptable. Could you share the link to the complete document?

@emeryberger Truly difficult times ahead for intro CS instructors :(

@jean72human I liked it better than Jekyll. Pretty easy to set up too. Plus lots of themes available.

@jean72human initially jekyll and later hugo

@psc Deep networks are just one parameterization, right? If you choose another parameterization, you could (potentially) update using gradient rules established in reinforcement learning (eg. updating mu, sigma through policy gradient rule assuming a gaussian policy w/ parameters mu, sigma). I don't know if it's popular in real world though. I think that some early IRL work also used non deep network parameterizations for reward functions.

@solalnathan I see, thanks for the link!

@psc I was gonna say moderate quality, moderate scores and moderate confidence 😂 😂

@solalnathan Is there any description of the battle transitions? For example, given two Pokemons where the user Pokemon makes a certain move, how are the new HPs calculated?

@solalnathan That sounds pretty cool, but is it a challenging environment?