Small 2D games with computer-controlled opponents trained using reinforcement learning.
You can play a demo in the browser here.
The goal of this project is to explore simple but fun game mechanics, machine learning for games, and deploying machine learning in the browser.
The main library is written in Python. Currently, a tag game and associated gymnasium environment for reinforcement learning is implemented. Training is done with Stable Baselines 3.
Note that Python 3.9 or greater is required. Installation is done with poetry:
git clone git@github.com:adamheins/shadows
cd shadows
poetry install
Play the game:
poetry run scripts/games/tag.py
If you have trained some models (more on this below), you can supply models for the computer-controlled agent to use when "it" and "not it" (recall this is a game of tag), using the options:
poetry run scripts/games/tag.py --it-model <path_to_it_model.zip> \
--not-it-model <path_to_not_it_model.zip>
If one or both are missing, then a default hand-crafted policy for the agent is used.
As mentioned above, models for when the computer agent is "it" and "not it" are
trained separately. This library provides the corresponding gymnasium
environments TagIt-v0 and TagNotIt-v0, respectively. To train, use
scripts/learn/train.py. For example, to train a soft actor-critic model on
the TagIt-v0 env for 1,000,000 timesteps and log the results:
poetry run scripts/learn/train.py TagIt-v0 -L logs --algo sac -n 1_000_000
A limited Javascript-version of the library is also implemented here.
Models trained using Python can be exported to the portable
ONNX format using scripts/learn/export_onnx_model.py.
MIT