This is the source code repository for the OMNI: Open-endedness via Models of human Notions of Interestingness paper. OMNI utilizes large (language) models as a model of interestingness, because they already internalize human concepts of interestingness from training on vast amounts of human-generated data. This repository implements OMNI on a procedurally generated 2D gridworld gomain Crafter.

• crafter/ folder contains Crafter environment code
• envs/ folder contains different environment settings presented in the paper
  • env_<tasks in env>_<method>.py scripts
  • tasks in env can be: tr (repetitive), trc (repetitive and compounds), trs (repetitive and synonyms)
  • methods can be: uni (uniform sampling), lp (learning progress), omni (OMNI), omoi (oracle)
• moi_saved/ folder contains the processed data from cached GPT-3 predictions
• evaluate.py script is used for visualizations and evaluations of a trained agent
• generate_plots.py script is used to generate the plots presented in the paper
• model.py specifies the model architecture
• train.py script is used to train the reinforcement learning agent

Clone the repository with git clone <repo_url> && cd omni_code.
Create python virtual environment python3 -m venv venv.
Activate python virtual environment source venv/bin/activate.
Install dependencies pip install -r requirements.txt.

Run train.py script with the necessary args:

python train.py --model <model_name> --env <env_name>

For example, in the repetitive task setting
Uniform: python train.py --model tr_uni-1 --env tr_uni --seed 1
LP: python train.py --model tr_lp-1 --env tr_lp --seed 1
OMNI: python train.py --model tr_omni-1 --env tr_omni --seed 1

Run evaluate.py script with the necessary args:

python evaluate.py --model <model_name> --env <env_name>

Run crafter/run_gui.py script with the necessary args:

python crafter/run_gui.py --env <env_name>

This codebase draws inspiration from the following codebases:

• Crafter
• torch-ac

@article{zhang2023omni,
  title={OMNI: Open-endedness via Models of human Notions of Interestingness},
  author={Jenny Zhang and Joel Lehman and Kenneth Stanley and Jeff Clune},
  year={2023},
  journal={arXiv preprint arXiv:2306.01711},
}