Recursive Skip-Step Planning (RSP)

This repository is the official implementation of Recursive Skip-step Planning in JAX, attached with the paper entitled Are Expressive Models Truly Necessary for Offline RL?.

Getting Started

Prerequisites

1. Install JAX

2. Install W&B and log in to your account to view metrics

3. Install the required dependencies:

pip install -r requirements.txt

Running experiment sets

To reproduce the D4RL benchmark results:

# Adroit
python exp_launcher.py include=experiment_conf/adroit.yaml
# AntMaze
python exp_launcher.py include=experiment_conf/antmaze.yaml
# MuJoCo
python exp_launcher.py include=experiment_conf/mujoco.yaml
# Franka Kitchen
python exp_launcher.py include=experiment_conf/kitchen.yaml

Metrics are uploaded to W&B. The final performance is also stored in the metrics folder. To consolidate into a markdown report, run the following command:

python view_metrics.py

This report will be saved in report.md.

Note: The experiment launcher will automatically allocate all idle GPUs on your machine and run experiments in parallel.

License

Apache License 2.0

Citation

@inproceedings{
          wang-niu2024rsp,
          title={Are Expressive Models Truly Necessary for Offline {RL}?},
          author={Guan Wang and Haoyi Niu and Jianxiong Li and Li Jiang and Jianming HU and Xianyuan Zhan},
          booktitle={NeurIPS 2024 Workshop on Open-World Agents},
          year={2024},
          url={https://openreview.net/forum?id=19KvVggjVr}
}