Nethack learning environment

[fph]

I just thought I'd write this post since there is a thread on the new borg and maybe not all users have heard of this fun recent research topic.

In the past few years there have been research efforts in the machine learning community to design artificial intelligences that play roguelikes; the standard (very challenging) testbed is Nethack. A first pointer is https://github.com/facebookresearch/nle . As far as I understand, they work by translating the game 'screen' into a textual description with tools like https://github.com/ngoodger/nle-language-wrapper , and then training a generic language-based AI Ă* la ChatGPT on it, via reinforcement learning.

As far as I know no AI has managed to ascend in Nethack yet. This report from a 2021 competition https://nethackchallenge.com/report.html mentions descending over 20 levels deep and achieving scores of over 100k points, though.

[Estie]

Very interesting. Do they get to know that dipping for Excalibur is a thing or do they have to learn everything from scratch ? I suspect the former, otherwise I dont see how can expert system can win the tournament.

[fph]

I think they are unspoiled. From what i understand, everything they know about the game comes from the 'language wrapper', which is essentially a textual description of what is shown on the screen.

EDIT: no, sorry, it seems that they know more; on this tutorial https://colab.research.google.com/dr...o=NEyq83OUBfKA they state that they use a database with a few million human-played games for their training.