Install dependencies
sudo apt-get update
sudo apt-get install -y python3-dev build-essential libgl1-mesa-glx libglib2.0-0
sudo apt-get install -y '^libxcb.*-dev' libx11-xcb-dev libglu1-mesa-dev libxrender-dev libxi-dev libxkbcommon-dev libxkbcommon-x11-dev libsm6 libxext6 libxrender-dev
(Optional) Build GRPC Emulator
mkdir -p cmake/build
cd cmake/build
cmake ../../
make
Create & activate venv
python3 -m venv venv
. venv/bin/activate
Install requirements & RetroGym
pip install -r requirements.txt
python setup.py install
You can customize an experiment by writing a config file. Refer to data folder for example config.
-
coreis the emulator library for Libretro to use when loading a ROM. -
romis the game ROM we use for experiments. -
stateis the game state to load when the game resets. -
actionsis the list of inputs allowed throughout the experiment. -
observationsis the list of variables we feed into the reinforcement model. The values are parsed directly from RAM according toaddress; to extract the values you need you need to refer to a RAM map. -
rewardsreads variables fromobservations, and evaluates rewards based on them. Currently available values forconditionarenew,duplicate,changed,unchangd, anddefault. -
endingsdetermines when to reset the environment based onobservations. Currently available available values forconditionarematchandtimeout.
To randomize the inputs, run the following command:
python src/main.py --config data/rival_battle.yaml --model random
To train a PPO model (implemented by Stable Baselines3), run:
python src/main.py --config data/rival_battle.yaml --model PPO --model-file models/rival --train
You may test the trained model by removing --train
python src/main.py --config data/rival_battle.yaml --model PPO --model-file models/rival
You should see something along the GIF below:
This project has gotten so far thanks to:
https://github.com/libretro/RetroArch
https://github.com/libretro/libretro-2048
https://github.com/Farama-Foundation/Gymnasium
https://github.com/openai/retro