This project implements novel team-based variants of card games, and presents a variety of deep reinforcement learning solutions that achieve good co-operative performance against differing levels of opposition.
The main contributions of this project lie in its focus on co-operative performance in card games. Novel reinforcement learning environments are offered for Uno and Leduc Hold'em which can be easily extended to other co-operative environments. A number of state-of-the-art reinforcement learning algorithms are also implemented.
This project is heavily based on RLCard, an existing open source Python card game reinforcement learning toolkit.
Existing RLCard codebase as well as this project's extensions to it exist in RLCard/. The basic structure of the codebase is as follows:
- RLCard/rlcard/models: Reinforcement Learning models using various algorithms are stored in this directory.
- RLCard/rlcard/agents: Reinforcement Learning agents for various algorithms are stored in this directory.
- RLCard/rlcard/envs: Reinforcement Learning environments for various card games are stored in this directory.
- RLCard/rlcard/games: Game logic for various card games are stored in this directory.
- RLCard/examples: Example Python scripts to use the RLCard library.
To demonstrate contributions of this project, examples/ contains a number of interactive Python scripts which run games against pre-trained agents using the solutions presented in this project. Example training notebooks are also available.
Finally, this project was a created as part of my BSc Computer Science Undergraduate dissertation. The accompanying research paper that outlines motivations, methodology, results and discussion is available in paper.pdf.
For installation and usage of the RLCard library please refer to the RLCard documentation.