kip93 / 3T-VI

A Deep Learning powered Virtual Intelligence that learns to play Tic-Tac-Toe

3T-VI, a Machine Learning powered Virtual Intelligence that learns to play Tic-Tac-Toe

The Tic-Tac-Toe Virtual Intelligence (or 3T-VI for short) aims to learn how to play by a process of trial and error, having minimal feedback. For this a small Deep Q Network has been built using Keras with a TensorFlow backend.

How does it work?

At the beginning 3T-VI plays pseudo-randomly in order to gather some initial information to work with. Then it uses a process of "exploration", trying out unused approaches, in an attempt to follow and weight as many paths as possible. After a while these random choices start to disappear, in order to "exploit" the more successful options, and avoiding the losing paths. Eventually it stops exploring altogether, sticking to the exploitation of all the gathered information choosing exclusively the best possible moves. This process is known as "Exploration and Exploitation" (D'oh).

Usage

The model itself can be found in the model.py file, and in there is everything needed to get it up and running, but a user interface has been provided as well, in the files view.py and main.py, containing the user interface and a controller between models and views, respectively. To run the user interface, simply run the main.py file.

In the given user interface, the approach taken to train 3T-VI was to put it against another 3T-VI. Why? Well, first, because this allows it to train against an equally good opponent, having at first random moves to help with the exploration phase of the training, and later they both play exploiting good moves, forcing each other to best its opponent. But most importantly:

Finally, for an added bonus, the provided user interface has been enhanced with some lame memes, for your entertainment.

Dependencies

The following lines list all dependencies and the oldest version they where tested against, along with instructions on how to install them.

Python (3.6+)

For windows, I recommend to download Anaconda. To install the rest of the dependencies use the generated Anaconda Prompt, and use conda instead of pip whenever possible.

For more details see Anaconda package list and installation guide

For Linux and OSx, install via your OS' package manager. For example in Ubuntu:

$ sudo apt install python3.6

Numpy (2.6+)

$ sudo pip install -U numpy

For more details, see SciPy installation guide.

TensorFlow (1.2+)

$ sudo pip install -U tensorflow

or

$ sudo pip install -U tensorflow-gpu

For more details, see TensorFlow installation guide.

h5py (2.7+)

$ sudo pip install -U h5py

For more details, see h5py installation guide.

Keras (2.0+)

$ sudo pip install -U keras

For more details, see Keras installation guide.

PyQt (5.6+)

Not required by the model itself, but by the user interface

$ sudo pip install -U pyqt5

For more details, see PyQt5 installation guide.