Dominion + Reinforcement Learning

This was originally a final project developed for CS238: Decision Making Under Uncertainty, Autumn 2019, by @garrickf and @andersonbcdefg. Read our paper, "Playing 'Dominion' with Deep Reinforcement Learning," here!

Presents a game simulator for the deck-building game Dominion (built in Python) and a deep reinforcement learning algorithm which develops strategies for the game (written in Keras/Python).

Setting up dependencies

Some libraries like numpy may need installation on your machine; these dependencies are held in requirements.txt. One can use venv in Python 3 to set up a virtual environment like so:

python3 -m venv env
source venv/bin/activate
pip install -r requirements.txt

venv is just the name of the environment and can be substituted for anything you want, although only venv is ignored in the git configuation for the repo. Be careful not to commit locally installed libraries!

Running experiments

To begin running experiments, the easiest thing to do is to mock one up in interactive mode. Run:

python learn_dominion.py -i

and you'll be taken through a series of questions to set up your experiment and the variables/hyperparameters therein. Interactive mode will tell you how to run those experiments again using just the CLI; for those of you who want to skip ahead:

usage: learn_dominion.py [-h] [--name NAME] [-m, --message MESSAGE] [-i]
                         [--niters NITERS] [--testevery TESTEVERY]
                         [--levels LEVELS]

Run Dominon learning experiment.

optional arguments:
  -h, --help            show this help message and exit
  --name NAME           experiment name
  -m, --message MESSAGE
                        comments on experiment/description
  -i, --interactive     setup experiment in interactive mode
  --niters NITERS       number of iterations to train for
  --testevery TESTEVERY
                        how often to test the policy
  --levels LEVELS       number of CPU levels to train against, tournament-
                        style

Playing a game of Dominion

To play Dominion against a computer agent, run

python game.py [weightfile]

where weightfile is the path to a weightfile that gets loaded into the neural network.

2-player human games, fixed and random policies coming soon...

More Documention coming soon...

Testing dominion

The dominion module implements unit tests to verify the functionality of the game and cards.

To run these tests, from the root directory of the repository, run:

python -m pytest

Tests are located in the /tests directory.

Style

Python code is formatted using black. isort is used for dependency sorting. Prettier is used for markdown formatting.