Official repository for "Direct Advantage Estimation"

We recommend using Python 3.8 with venv. Please make sure pip is up to date by running:

pip install -U pip

Install requirements:

pip install -r requirements.txt

To reproduce the results, run the following command:

python train.py --algo {algo} --hparam_file {hyperparameter_file} --envs {env} --threads {threads}

--algo: PPO (GAE) or CustomPPO (DAE)

--hparam_file: See ./params/ for the hyperparameters used in the paper, the files are named by {algo}_{network}.yml

--envs: Environment to train. For example, Pong, Breakout, etc. For MinAtar environments, please add the suffix -MinAtar-v0. (e.g., Breakout-MinAtar-v0)

--threads: Number of parallel threads for asynchronous environment steps

--logging: Save logs in ./logs/{env}/

--save_model: Save the trained model to ./logs/{env}/

To view the tensorboard logs, run

python -m tensorboard --logdir ./logs/

and open the displayed URL in a browser.

Please use the following BibTex entry.

@article{pan2022direct,
  title={Direct advantage estimation},
  author={Pan, Hsiao-Ru and G{\"u}rtler, Nico and Neitz, Alexander and Sch{\"o}lkopf, Bernhard},
  journal={Advances in Neural Information Processing Systems},
  volume={35},
  pages={11869--11880},
  year={2022}
}