Keras zero knowledge AI for gomoku and tic-tac-toe

Blogpost for the algorithm: http://blog.teamleadnet.com/2018/07/zero-knowledge-ai-for-tic-tac-toe-and.html

Prerequisites

Keras with a valid backend (eg, tensorflow)

pip install tensorflow
pip install keras

To install optimised Tensorflow backend, go to https://github.com/lakshayg/tensorflow-build

Training

The system will start playing randomly against itself with the predefined parameters.

For board size 6x6 with 5 in a row for winning, stopping after random only wins 2% of the games

./ai_gomoku.py train --boardsize 6 --winsize 5 --stopon 2 --decay 0 --lr 0.001 --layer1size 128 --layer2size 128 --epochs 2 --randomuntil 0 --updatebatchsize 10000

Training time ~5 hours without GPU.

To watch the progress, start

tensorboard --logdir=logs/

and head to http://localhost:6006

Playing agianst the model

The player starts with X. Input the moves in the format of X, Y where X is the row, Y is the column, starting from zero.

./ai_gomoku.py play --boardsize 6 --winsize 5 --filename gomoku_6_5.h5 

Eg,

|      |
|      |
|      |
|      |
|      |
|      |
Your move? (x, y) 3,3
|      |
|      |
|   0  |
|   X  |
|      |
|      |
...

Evaluating boards

Copy-paste boards to STDIN to evaluate it's score based on a model. Boards are separated with two empty lines. End with a CMD+D.

Sample boards are provided in eval_boards.txt

To indicate empty rows add a single space in that line.

./ai_gomoku.py eval --boardsize 6 --winsize 5 --filename gomoku_6_5.h5