Tetris AI using Deep Reinforcement Learning

This projects implements a Tetris AI using double deep Q-learning.

Requirements

• Current version of Python3 (I used Python 3.10)
• Jupyter Notebook/Lab
• The package manager pip

Use pip to install the dependencies:

pip install numpy pandas tqdm ipywidgets wandb pygame torch

Tested with Ubuntu 22.04.

Installation

Clone the project using git.

Extract the file replays.zip in the directory pro-replays so all the .json files lie directly in the pro-replays directory. This is needed since unzipped those files are nearly 250MB in size, while zipped only about 9MB.

Usage

Open the cloned folder in Jupyter. The main learning approach is in the file dqn.ipynb.

After training an agent with the dqn.ipynb notebook and saving it to the models directory, this agent can be used to render a run using the run-renderer.ipynb. You just have to change the name of the saved/loaded model accordingly in both notebooks.

A pretrained agent's weights can be found in models/good-cnn-1.pt and models/good-cnn-2.pt. By providing those files to the run-renderer.ipynb one can watch them play.

Folder Structure

• dqn.ipynb: the main file that implements the double deep Q-learning approach
• run-renderer.ipynb: uses learned model weights to simulate a game played by the agent and renders it in a graphical interface (using pygame)
• tetris/: contains the implementation of the game
• gui/: contains the code for the GUI renderer
• util/: helpers (currently only implementation for decay-functions)
• models/: contains the learned weights for selected runs

Demo

To get to the better quality and longer youtube video version, click on the GIFs.

CNN

Note: The genetic approach version can currently be found in the genetic branch of this repository.

Genetic Singles

Genetic Quads