This project was developed for the subject 'Laboratório de IACD' by the students André Sousa, António Cardoso, Antónia Brito and Paulo Silva.
In this project, our work which takes focus on recreating an adversarial model inspirated on DeepMind's AlphaZero capable of learning to play the games of Go and Ataxx.
To achieve this, we created a machine learning model based on the AlphaZero aproach, making use of a Convolutional Neural Network and a Monte Carlo Tree Search algorythm.
On top of that we implemented some features of our own, such as:
- Multiple games can be played in parallel
- A model that can play on different sized boards
- Data Augmentation by applying transformations to board states
- Introduction of noise to mitigate overfitting
The final work is distribuited on theese files:
- aMCTS_parallel.py: which implements the MCTS
- az_parallel2.py: containing the AlphaZero class
- ataxx.py and go.py: that implement the rules of the games
- graphics.py: to display the pygame interface
- play_AI.py: to play against a trained model
- ataxx4x4.py, ataxx5x5.py, ataxx6x6.py, ataxxflex.py, go7x7.py and go9x9.py: to train each model
If you wish to play against your models, you can do it by changing the game characteristics, the model characteristics and the model path in play_AI.py and running it.
README.md by André Sousa