An Application of Modern Reinforcement Learning Algorithms to The Card Game Cego

General Information

• supervised by: Prof. Dr. Maja Temerinac-Ott

• cosupervised by: Prof. Jirka Dell´Oro-Friedl

• submitted on: 31.08.2022

• submitted by:

  • Philipp Oeschger
  • matriculation number: 268388
  • Bregstraße 90
  • 78120 Furtwangen im Schwarzwald
  • [email protected]

• Link to Master Thesis

• The folder src contains the source code to this thesis

About this Repository

This repository contains the full source code of the thesis.

The repository is a fork of RLCard and, therefore, the src/rl_env folder contains external source code. The following files / folders in src/rl_env do contain self created code:

• src/rl_env/rlcard/games/cego
• src/rl_env/rlcard/envs/cego.py
• src/rl_env/rlcard/agents/human_agents/cego_human_agent.py
• src/rl_env/results
• src/rl_env/python_scripts

Description of source code stucture the structure:

Class Diagramm of Game Implementation:

File Stucture of src folder:

Deep-AI-Service

The API that makes the AI-models available.

• this is a seperate service that requires a seperate setup
• more details here

RL-Env

• rlcard/games/cego: The implementation of the game logic.
  • jsondata: contains card encoding, high cards and low cards list.
  • testing: contains script for testing the environment
  • utility: contains utility function modules for:
    • evaluations
    • game implementation
    • custom logging
    • random search Hyperparameter optimization
    • training
  • ... game classes described in class diagram
• rlcard/envs/cego.py: The environment class for the game.
• rlcard/agents/human_agents/cego_human_agent.py: An agent that serves as testing interface to play against AI models.
• results: contains various results, including
  • analysis results,
  • evaluations,
  • model training,
  • random search results.
• python_scripts: script for training, hyperparameter search and evaluation.
  • cego_random_search: random search scripts for dqn and nfsp.
  • final_model_training: contains,
    • evaluation of models
    • analysis of the game environment
    • scripts for final dmc, dqn, and nfsp training
  • interactive_cego_testing: contains a test script for playing against a model in the command line

Setting up the Environment

The following tools are needed to setup the environment:

• Python 3 (Python 3.9 was used)
• Virtualenv

Open rl_env_folder

cd src/rl_env

Create an environment

python3 -m venv venv

Activate Environment

MacOS:

source venv/bin/activate

Linux:

source venv/Scripts/activate

Windows:

.\venv\Scripts\activate

Install the Dependencies

pip install -r requirements.txt

pip3 install -e .

Run Tests for RL-Env:

python -m unittest discover