This repository contains the source code for the WFCRL multi-agent RL benchmark. The benchmark is done on the WFCRL environment suite.
All experiments are adapted from the CleanRL repository. Algorithms:
| Algorithm | File | Description |
|---|---|---|
| IPPO | algos/baseline_ippo.py |
See Yu et. al |
| MAPPO | algos/baseline_mappo.py |
See Yu et. al |
| QMIX | algos/baseline_qmix.py |
See Rashid et. al |
| IFAC | algos/ifac.py |
Simple online actor critic with Fourier Basis |
| IQN | algos/idqn.py |
Simple independent DQN |
Scripts with the windrose suffix train under Wind Scenario II. Other implement Wind Snecario I.
Install the dependencies:
pip install -r requirements
Launch an IPPO training experiment on the Dec_Ablaincourt_Floris environment:
python algos/baseline_ippo.py --seed 1 --env_id Dec_Ablaincourt_Floris --total_timesteps 1000000
Evaluate it on the on the Dec_Ablaincourt_Fastfarm environment:
mpiexec -n 1 python algos/eval.py --seed 0 --env_id Dec_Ablaincourt_Fastfarm --total_timesteps 10000 --pretrained_models path/to/run
Experiments for training and evaluation runs are in the scripts folder.
To track the experiment in Wandb, add your API key in an .env file at the root of the folder:
WANDB_API_KEY=you_api_key