KO | EN

GitLite

Minimalist Open Source Repository Search Engine

gitlite — search

#typescript #ai-agents #cli #open-source #ai #claude-code #codex #developer-tools #dsh-plugin #ai-agent #mcp #deepseek-harness

PEDA

★ 34 Open GitHub ↗

Scaling Pareto-Efficient Decision Making via Offline Multi-Objective RL, published in ICLR 2023

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# Scaling Pareto-Efficient Decision Making via Offline Multi-Objective RL (ICLR 2023)

[Website](https://baitingzbt.github.io/projects/iclr_2023_morl/) | [Poster](https://iclr.cc/media/PosterPDFs/ICLR%202023/11257.png?t=1680814838.1065722) | [OpenReview](https://openreview.net/forum?id=Ki4ocDm364)

Authors: [Baiting Zhu](https://baitingzbt.github.io/), [Meihua Dang](https://cs.stanford.edu/~mhdang/), [Aditya Grover](https://aditya-grover.github.io/)

## Setup
  ```
  git clone https://github.com/baitingzbt/PEDA.git
  cd PEDA
  conda env create -f environment.yml
  conda activate peda_env
  ```

## Data Download
This folder contain all dataset variants used in the paper experiments including ablation study. All variants: check "generate your own data" section below.
```
pip install gdown
gdown --folder https://drive.google.com/drive/folders/1wfd6BwAu-hNLC9uvsI1WPEOmPpLQVT9k?usp=sharing --output data
```
The "data" folder should be under "PEDA" e.g.: `PEDA/data/env/data_name.pkl`
## Training
First double-check your CUDA devices and data path in this shell script. Run the uniform experiments for all environments:
```
sh all_env_uniform.sh
```
Alternatively, here is an example for a single experiment:
```
python experiment.py --dir experiment_runs/uniform --env MO-HalfCheetah-v2 --data_mode _formal --concat_state_pref 1 --concat_rtg_pref 0 --concat_act_pref 0 --mo_rtg True --seed 1 --dataset expert_uniform --model_type rvs --num_steps_per_iter 200000 --max_iters 2
```

## Generate Your Own Data (WIP)
Due to storage limit, we cannot easily open-source all data variants. Please check the source code to collect data. First download ckpts from  [https://drive.google.com/file/d/1_-B0UAt57JV-Jc1aBYEasuEgmR7nouaf/view?usp=sharing](https://drive.google.com/file/d/1_-B0UAt57JV-Jc1aBYEasuEgmR7nouaf/view?usp=sharing). Unzip, rename folder to `Precomputed_Result`, and move this folder under `data_generation`.
```
# DOWNLOAD, UPZIP, RENAME, MOVE

# USE AFTER MANUAL SETUP
cd data_generation
sh collect_all.sh
```
Note 1: We use randomly-initialized environments which is different from behavioral policy paper. This helps to diversify trajectories.

Note 2: Model ckpts are stored under `PEDA/data_generation/Precomputed_Results`. All were kindly provided by the authors of behavioral policy paper, except that we trained Hopper-v3 ourselves.
## Citation
If you use this repo, please cite:
```
@inproceedings{
    zhu2023paretoefficient,
    title     = {Scaling Pareto-Efficient Decision Making via Offline Multi-Objective RL},
    author    = {Baiting Zhu and Meihua Dang and Aditya Grover},
    booktitle = {International Conference on Learning Representations},
    year      = {2023},
    url       = {https://openreview.net/forum?id=Ki4ocDm364}
}
```

PEDA

Explore Similar Repositories

Was this content helpful?

Recent Feedback

Download README