Solving the UNO Card Game using Monte-Carlo Reinforcement Learning Model

Abebual Zerihun Demilew and Rachel Fisher

---

Image from Kohls.com

---

In this project we developed a Monte-Carlo Reinforcment Learning Agent to play an optimal strategy to win a simplified scope of UNO Card Game. The project is stractured as follows:

1. Game Environment
2. Monte-Carlo Reinforcment Learning Agent
3. Model Training and Performance
4. Experments

1. Game Environment

In environment.py class objects for card, deck, player, turn and game are defined. In main.ipynb, the classes are imported as module, to run simulations. The state_action_reward.py includes the state, action, and reward class objects.

2. Monte-Carlo Reinforcment Learning Agent

The monte-carlo method reinforcment learning algorithm is defined in agents.py module.

3. Model Training and Performance

Results from model training included in the project dataset folder - [Rachel Fisher]-[Abebual Demilew]--dataset.zip. Monte-carlo RL model outputs such as q-tables, q-values coverage, win-rates, turns, state-action pairs and visits are included in the dataset folder.

Model performance results from best performing monte-carlo RL Agent:

Frequency of playable cards available:

4. Experments

Additional experments were conducted to:

Understand if there is a first player advantage.

Understand the exploration vs exploitation tradeoff.

Understand if any card has a playable advantage.

Python Version: 3.7
Packages: pandas, numpy, random, itertools, time, tqdm, sys, os, matplotlib, seaborn, ipywidgets