Pokemon Showdown AI

Reinforcement learning project for Pokemon Showdown. Currently only supports the Gen-4 random battle format.

This project has three parts:

PsBot framework for creating a general Pokemon Showdown bot and setting up the battle interface.
Battle state tracker and PS protocol parser.
Neural network training script.

2022-08-14_demo.mp4

Me (left) vs a model (right) that was trained over ~16k games against itself

Build Instructions

Make sure you have at least Node v16 (LTS) and Miniconda 3 installed. Should work on Linux and likely also Windows WSL2.

# Download the repository.
git clone https://github.com/taylorhansen/pokemonshowdown-ai
cd pokemonshowdown-ai

# Checkout submodules.
git submodule init
git submodule update

# Setup Python/TensorFlow.
conda env create --name psai --file environment.yml
conda activate psai

# If using GPU, ensure cuDNN gets loaded correctly.
# Copied from https://www.tensorflow.org/install/pip.
mkdir -p $CONDA_PREFIX/etc/conda/activate.d
echo 'CUDNN_PATH=$(dirname $(python -c "import nvidia.cudnn;print(nvidia.cudnn.__file__)"))' >> $CONDA_PREFIX/etc/conda/activate.d/env_vars.sh
echo 'export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$CONDA_PREFIX/lib/:$CUDNN_PATH/lib' >> $CONDA_PREFIX/etc/conda/activate.d/env_vars.sh

# Setup TS.
npm install
npm run build

Testing

# Run formatter.
npm run format
isort src test
black src test

# Run linter.
npm run lint
pylint src test
mypy src test

# Run tests.
npm test
python -m test.unit

Training

# Edit hyperparmeters as needed.
cp config/train_example.yml config/train.yml

python -m src.py.train

Trains the neural network through self-play. This requires a powerful computer and/or GPU, and may take several hours depending on how it's configured.

Training logs are saved to ./experiments/ by default.

Metrics such as loss and evaluation scores can be viewed using TensorBoard.

pip install tensorboard
tensorboard --logdir experiments

Running

# Edit config as needed.
cp config/psbot_example.yml config/psbot.yml

npm run psbot

Connects to the PS server specified in the config and starts accepting battle challenges in the gen4randombattle format, which is the only format that this project supports for now. By default it loads the model from ./experiments/train/model (assuming a training run was completed) and connects to a locally-hosted PS instance (see guide on how to set one up). This allows the trained model to take on human challengers or any other outside bots.

License