Applying Reinforcement Learning to the Card Game Wizard

This repo contains the implementation for an agent learning to play the card game wizard through Reinforcement Learning. In the current implementation a version of Proximal Policy Optimization was used. However, it is trivial to use other algorithms for learning with the current setup.

This work was created during the course Neural Information Processing Projects at TU Berlin in the spring term 2019. Supervisor was Vaios Laschos

Setup

Install dependencies:
pip install -r requirements.txt

Repo Structure

• agents --> contains the implementation for the rl agents and rule based agent. tf-agents library is used
  • contains implementation for predictors, featurizers
• evaluation --> saves models and logs during evaluation process
• game engine --> contains implementation of game logic
• tests --> contains code for training and evaluation the agent
• wizard_site --> contains implementation for GUI. Used for evaluating agents against human players

Usage

To train the agent: \
$ python3 tests/test_train_and_evaluate.py train_vs_old_self

To evaluate the agent: \
$ python3 tests/test_train_and_evaluate.py evaluate

The above commands will also start a tensorboard session on port 6006.

Also see the comments and docstrings in tests/test_train_and_evaluate.py

To start the GUI: \
From the root directory run: python wizard_site/manage.py runserver

Paper

For more detailed information on the implementation and results check the paper and poster in /paper

Contributers

Jonas Dippel, Callum Waters, Til Jasper Ullrich, Kai Jeggle