Othello AI

Play Mode

You can play against AI

$ python src/game.py model/slpn.best_accuracy.npz --first

(Demo)

Data Preparation

Download records of an Othello game from the following URL

https://www.skatgame.net/mburo/ggs/game-archive/Othello/

Cleanse the data

$ python scripts/cleanse_data.py data/hoge data/cleansed.txt

Extract board and move pairs.

$ python scripts/restore_from_data.py data/cleansed.txt data/restored.txt

You will see these lines in restored.txt.

> 0000000000000000000100000001100000010000000000000000000000000000 0000000000000000000000000000000000001000000000000000000000000000 W 18
> 0000000000000000000100000011100000010000000000000000000000000000 0000000000000000001000000000000000001000000000000000000000000000 W 20
> ...

Each column shows

where black discs are placed (expressed by 64-bits number)
where white discs are placed (expressed by 64-bits number)
the color of a player who makes a next move
next move (expressed by an integer from -1 to 63)

This script only extracts the winner's actions to select from more than one valid moves.

Create Dataset

Run the following command, and train/validation data will be created in the same directory as an input file.

$ python scripts/create_dataset.py data/restored.txt --train-size {train-size} --valid-size {valid-size}

You may fail to create dataset if too large number is specified.

Environment

$ pip install pipenv --user
$ pipenv install
$ pipenv shell

Model

In progress ...

Now supervised-learning policy network is available

$ python src/train.py config/sample.json [**kwargs]

Result

data	size
Train	10,000,000 pairs
Valid	624,169 pairs

(no duplicates between train and valid)

model	max validation accuracy
supervised-learning policy network	0.632