Achievement Distillation

This is the code for the paper Discovering Hierarchical Achievements in Reinforcement Learning via Contrastive Learning accepted to NeurIPS 2023.

Installation

conda create --name ad-crafter python=3.10
conda activate ad-crafter

pip install --upgrade "setuptools==65.7.0" "wheel==0.38.4"
pip install -r requirements.txt
pip install -e .

Usage

PPO (baseline)

python train.py --exp_name ppo --log_stats

PPO + Achievement Distillation (ours)

python train.py --exp_name ppo_ad --log_stats

Citation

If you find this code useful, please cite this work.

@inproceedings{moon2023ad,
    title={Discovering Hierarchical Achievements in Reinforcement Learning via Contrastive Learning},
    author={Seungyong Moon and Junyoung Yeom and Bumsoo Park and Hyun Oh Song},
    booktitle={Neural Information Processing Systems},
    year={2023}
}