README.md

Q-ensemble for Offline RL: Don't Scale the Ensemble, Scale the Batch Size

This repository contains the official implementation of Q-ensemble for Offline RL: Don't Scale the Ensemble, Scale the Batch Size.

Dependencies & Docker setup

To set up python environment (with tool on your taste, in our research we use conda and python 3.8), just install all requirements:

python install -r requirements.txt

However, with such setup, you would also need to install mujoco210 binaries by hand. Sometimes this is not straightforward, but we used this recipe:

mkdir -p /root/.mujoco \
    && wget https://mujoco.org/download/mujoco210-linux-x86_64.tar.gz -O mujoco.tar.gz \
    && tar -xf mujoco.tar.gz -C /root/.mujoco \
    && rm mujoco.tar.gz
export LD_LIBRARY_PATH=/root/.mujoco/mujoco210/bin:${LD_LIBRARY_PATH}

You may also need to install additional dependencies for mujoco_py. We recommend following the official guide from mujoco_py.

Docker

We also provide a simpler way, with a dockerfile that is already set up to work, all you have to do is build and run it :)

docker build -t lb_sac .

To run, mount current directory:

docker run -it \
    --gpus=all \
    --rm \
    --volume "<PATH_TO_THE_REPO>/lb-sac:/workspace/lb-sac" \
    --name lb_sac \
    lb_sac bash

How to reproduce experiments

Configs are stored in configs/<algo_name>/<task_type>. All available hyperparatemers are listed in train.py. For example, to start EDAC training on halfcheetah-medium-v2:

python train.py \
    --config_path="configs/edac/halfcheetah/halfcheetah_medium.yaml" \
    --device=cuda

You can also use our simple tool for sweeps and multiseed training (same can be done with wandb sweep):

python sweep.py \
    --command "python train.py" \
    --configs "configs/edac/*" \
    --num_seeds 4 \
    --num_gpus 8

This will run each config for 4 seeds on all available gpus in the queue (8 at max).

Citing

If you use this code for your research, please consider citing the paper:

@article{nikulin2022q,
  title={Q-Ensemble for Offline RL: Don't Scale the Ensemble, Scale the Batch Size},
  author={Nikulin, Alexander and Kurenkov, Vladislav and Tarasov, Denis and Akimov, Dmitry and Kolesnikov, Sergey},
  journal={arXiv preprint arXiv:2211.11092},
  year={2022}
}