Here are a few multi-agent games that can be learned using RL.
pursuit
, use the command python3 main_run.py pursuit
.waterworld
, use the command python3 main_run.py waterworld
.multiwalker
, use the command python3 main_run.py multiwalker
.This repo uses the Reinforcement Library toolkit RLlib
from Ray. The specific wheels Ray
package that are needed are included in this repo (there have been breaking changes). Once the wheel is downloaded, install it as pip install -U path/to/wheel.whl
. Install the other required packages using the command pip install -r requirements.txt
. Required Python version is 3.7.6
.
python3 parameterSharingPursuit.py RLmethod
to train an RL method RLmethod
(e.g. PPO) on pursuit
.python3 parameterSharingMultiwalker.py RLmethod
to train an RL method RLmethod
(e.g. PPO) on multiwalker
.python3 parameterSharingWaterworld.py RLmethod
to train an RL method RLmethod
(e.g. PPO) on waterworld
.Use play_pursuit.py
, play_waterworld.py
and play_multiwalker.py
to re-play the games using the learned RL policies. Simply change the checkpoint_path
variable and make sure params.pkl
is in the correct place, relative to the former.