AI611 Nethack Project

Project Repository for Nethack Challenge .

AI611: Deep Reinforcement Learning, KAIST GSAI 2022SP

How to train the agent

First, run the docker container:

$ ./docker/run.sh

Then, install additional dependencies within the docker container:

$ python3 -m pip install stable_baselines tensorboard

(Note that this step is needed, since we do not build a custom docker image, but use the image provided by fairnle/challenge:dev.)

Afterwards, you can freely train our agent:

$ python3 train_pop_art_agent.py

Optional tensorboard visualization:

$ tensorboard --logdir /tmp/nethack --host 0.0.0.0 --port 6006

Afterwards, open localhost:6006 on your browser to see the logged outputs.

How to record the video

After training the agent, you can record the video as follows:

Link to pretrained checkpoint

$ python3 record.video.py ${CKPT_FILE} # e.g. '/tmp/nethack/run-022/log/nh-pa-last.pt'
$ python3 -m nle.scripts.ttyplay /tmp/record/nle.PPPPP.I.ttyrec.bz2

Where PPPPP is the process ID and I is the environment index. In general, you'd want to set PPPPP to the latest one, and I = 0.

Collaborators

• Yoonyoung(Jamie) Cho Github, Website
• Minchan Kim Github
• Heesang Jo Github