This repo depends on pytorch (version 1.9.0) and mujoco (version 200). I am running python 3.7 on my system.
Other requirements can be installed via pip by running:
pip install -r requirements.txt
Data must be saved into a replay buffer in the data/ directory by running e.g. python d4rl_to_replay.py --name halfcheetah-medium-v2 for whichever D4RL dataset you want data from.
The train loop is in train.py which can capture all algorithms in the paper by varying parameters as described below.
Important: to run the training files, you need to include the path from the root to the onestep-rl directory on your machine as the path variable in the config/train.yaml file, e.g. path: /path/to/onestep-rl.
Config files with all the relevant hyperparameters can be found in the config/ directory.
To get the one-step algorithm from the paper we set the following parameters of the training loop in config/train.yaml:
beta_steps: 5e5
steps: 1
qs_teps: 2e6
pi_steps: 1e5
To get the iterative algorithms we load the pre-trained beta and q estimators used by the one-step algorithm and then run
beta_steps: 0
steps: 1e5
q_steps: 2
pi_steps: 1
For the multi-step algorithms we load the pre-trained beta and q estimators used by the one-step algorithm and then run
beta_steps: 0
steps: 5
q_steps: 2e5
pi_steps: 2e4
All the figures from the paper along with the notebooks that generated them are in the figures/ directory. Due to space constraints, the data and log files needed to generate the figures are not included.
If you use this repo in you research, please cite the paper as follows
@article{brandfonbrener2021offline,
title={Offline RL Without Off-Policy Evaluation},
author={Brandfonbrener, David and Whitney, William F and Ranganath, Rajesh and Bruna, Joan},
journal={arXiv preprint arXiv:2106.08909},
year={2021}
}