This project is a simple and extendable AI demonstration. The code is written in Python.
The GUI is built on TK, and defensive programming is mostly ignored.
To test the application, simply start main.py .
You can modify main.py to change the way the application behaves. Sorry, I was too lazy for command line support.
Currently, there are two supported modes:
- Training
- Execution
ai.abstract_ai is the base class for any AI code to play tic tac toe. You can add new AI implementations by inheriting a new class from ai.abstract_ai.
The abstract class has methods for:
- Training itself & saving train data
- Loading train data
- Playing with humans & learning
The debut release of this project includes a sample & simple Q-Learning implementation; which conducts an unsupervised reinforcement learning effort.
If you don't know anything about Q-Learning, you can get more info here.
In training mode, the program will play against itself with 80% random moves and 20% of learned moves during the training session. When a game is finished, the moves will be evaluated. The number of games to be played can be modified in ai.qlearning. The initial number is 10K.
For the winner, the last move is worth n points, and former moves are worth n-1, n-2, ... points.
For the loser, the last move is worth -n points, and former moves are worth -(n-1), -(n-2), ... points.
Evaluation of games are stored in the qlearning_memory.txt file. A sample dump file is included. As the code plays against humans, it keeps the file updated.
AI is also able to do board mirroring. If it can't find a good move in the database (text file), it will rotate & flip the board and seek similar patterns.
In a semi-supervised training approach, I could have given the ability to foresee obvious good / bad moves.
It is also possible to re-play lost games by changing one move at a time until the game is won. That would improve the learning effectiveness of the code.
The results are stored in a text file. For larger projects, a database could have been a more suitable choice.