AnukritiSinghh / MRN

Official codebase for "Meta-Reward-Net: Implicitly Differentiable Reward Learning for Preference-based Reinforcement Learning" (NeurIPS 2022).

Meta-Reward-Net

Official implementation of the NeurIPS 2022 paper: "Meta-Reward-Net: Implicitly Differentiable Reward Learning for Preference-based Reinforcement Learning". Our implementation is based on the official codebase of B-Pref and SURF.

How to install

Install MuJoCo 2.0

sudo apt update
sudo apt install unzip gcc libosmesa6-dev libgl1-mesa-glx libglfw3 patchelf libegl1 libopengl0
sudo ln -s /usr/lib/x86_64-linux-gnu/libGL.so.1 /usr/lib/x86_64-linux-gnu/libGL.so
wget https://www.roboti.us/download/mujoco200_linux.zip -P /tmp
unzip /tmp/mujoco200_linux.zip -d ~/.mujoco
wget https://www.roboti.us/file/mjkey.txt -P /tmp
mv /tmp/mjkey.txt ~/.mujoco/

Install dependencies

conda env create -f conda_env.yml
conda activate mrn
pip install -e .[docs,tests,extra]
cd custom_dmcontrol
pip install -e .
cd ../custom_dmc2gym
pip install -e .
cd ..
pip install git+https://github.com/rlworkgroup/metaworld.git@master#egg=metaworld
pip install pybullet termcolor

How to run

DeepMind Control Suite (Walker)

MRN

python train_MRN.py env=walker_walk seed=12345 agent.params.actor_lr=0.0005 agent.params.critic_lr=0.0005 num_unsup_steps=9000 num_train_steps=1000000 num_interact=20000 max_feedback=100 reward_batch=10 reward_update=50 feed_type=1 num_meta_steps=1000

PEBBLE

python train_PEBBLE.py env=walker_walk seed=12345 agent.params.actor_lr=0.0005 agent.params.critic_lr=0.0005 num_unsup_steps=9000 num_train_steps=1000000 num_interact=20000 max_feedback=100 reward_batch=10 reward_update=50 feed_type=1

SURF

python train_PEBBLE_semi_dataaug.py env=walker_walk seed=12345 agent.params.actor_lr=0.0005 agent.params.critic_lr=0.0005 num_unsup_steps=9000 num_train_steps=1000000 num_interact=20000 max_feedback=100 reward_batch=10 reward_update=1000 inv_label_ratio=100 feed_type=1 threshold_u=0.99 mu=4

Meta-world (Door Open)

MRN

python train_MRN.py env=metaworld_door-open-v2 seed=12345 agent.params.actor_lr=0.0003 agent.params.critic_lr=0.0003 num_unsup_steps=9000 num_train_steps=1000000 agent.params.batch_size=512 double_q_critic.params.hidden_dim=256 double_q_critic.params.hidden_depth=3 diag_gaussian_actor.params.hidden_dim=256 diag_gaussian_actor.params.hidden_depth=3 num_interact=5000 max_feedback=1000 reward_batch=10 reward_update=10 feed_type=1 num_meta_steps=10000

PEBBLE

python train_PEBBLE.py env=metaworld_door-open-v2 seed=12345 agent.params.actor_lr=0.0003 agent.params.critic_lr=0.0003 num_unsup_steps=9000 num_train_steps=1000000 agent.params.batch_size=512 double_q_critic.params.hidden_dim=256 double_q_critic.params.hidden_depth=3 diag_gaussian_actor.params.hidden_dim=256 diag_gaussian_actor.params.hidden_depth=3 num_interact=5000 max_feedback=1000 reward_batch=10 reward_update=10 feed_type=1

SURF

python train_PEBBLE_semi_dataaug.py env=metaworld_door-open-v2 seed=12345 agent.params.actor_lr=0.0003 agent.params.critic_lr=0.0003 num_unsup_steps=9000 num_train_steps=1000000 agent.params.batch_size=512 double_q_critic.params.hidden_dim=256 double_q_critic.params.hidden_depth=3 diag_gaussian_actor.params.hidden_dim=256 diag_gaussian_actor.params.hidden_depth=3 num_interact=5000 max_feedback=1000 reward_batch=10 reward_update=20 feed_type=1 inv_label_ratio=10 threshold_u=0.99 mu=4

Citation

@inproceedings{liu2022metarewardnet,
  title={Meta-Reward-Net: Implicitly Differentiable Reward Learning for Preference-based Reinforcement Learning},
  author={Runze Liu and Fengshuo Bai and Yali Du and Yaodong Yang},
  booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
  editor={Alice H. Oh and Alekh Agarwal and Danielle Belgrave and Kyunghyun Cho},
  year={2022}
}