Self-contained implementation of paper 'Beyond Non-Expert Demonstrations: Outcome-Driven Action Constraint for Offline Reinforcement Learning'

This is an experimental demo based on YangRui2015's RORL project.

1. Requirements

To install the required dependencies:

conda create -n cum python=3.7
conda activate cum
pip install --no-cache-dir -r requirements.txt

2. Usage

2.1 Training

python -m scripts.sac --env_name [ENVIRONMENT] --num_qs 10 --norm_input --load_config_type 'benchmark' --exp_prefix RORL

2.2 Evaluation

python -m scripts.sac --env_name [ENVIRONMENT] --num_qs 10 --norm_input --eval_no_training --load_path [model path] --exp_prefix eval_RORL