Diaster: Episodic Return Decomposition by Difference of Implicitly Assigned Sub-trajectory Reward

This is the code for the paper "Episodic Return Decomposition by Difference of Implicitly Assigned Sub-trajectory Reward".

Installation instructions

Install Python environment with:

conda create -n diaster python=3.9 -y
conda activate diaster
conda install pytorch cudatoolkit=11.3 -c pytorch -y
pip install -r ./requirements.txt

Run an experiment

python3 main.py --env-name=[Env name] 

The config files act as defaults for a task.

They are all located in config. --env-name refers to the config files in config/ including Hopper-v3, Walker2d-v3, Swimmer-v3, Humanoid-v3, HumanoidStandup-v2.

All results will be stored in the result folder.

For example, run Diaster on Hopper:

python main.py --env-name=Hopper-v3