fun-with-ml-tic-tac-toe-ai

The goal of this project was to create a Machine Learning based AI for Tic Tac Toe game.

Content:

• lib - game emulation libs
• ai - different AI implementations (random, Minimax, NeuralNetwork)
• ai-utils - utilities for neural network AI training
• ai-models - pretrained neural network for NN AI
• data - data gathered for NN AI training
• stats - utils to produce statistics about different AI combinations
• test - mocha tests for libs and AIs
• manual-tests - some aux tests used durng developement

Check out AIs

You can easily check out pretrained NN AI using utils in stats folder this way:

node stats/stats-rand-rand.js
node stats/stats-mm-mm.js
node stats/stats-mm-rand.js
node stats/stats-nn-nn.js
node stats/stats-rand-mm.js
node stats/stats-rand-rand.js
node stats/stats-mm-nn.js
node stats/stats-nn-mm.js
node stats/stats-nn-rand.js
node stats/stats-rand-nn.js

As you can see Random AI is lame but fast, Minimax AI is good but very slow and NN AI is quite good and quite fast.

Train your own NN AI

It is quite simple:

• clean up data and ai-models folders
• generate data for learning: node ai-utils/collect-minimax-debug-csv.js
• prepare data: ./ai-utils/prepare-datasets.sh
• create and train NN (here you can adjust training parameters): node ai-utils/create-and-train-nn-brain-js.js
• test NN: node ai-utils/test-ai-nn-brain-js.js

This project uses

• NodeJS (https://nodejs.org/en/)
• brain.js (https://github.com/harthur-org/brain.js)
• async (https://github.com/caolan/async)
• mocha (https://github.com/mochajs/mocha)

I used TensorFlow (https://www.tensorflow.org/) to figure out the shape of required NN - it is really fast!

Why callbacks and process.nextTick???

The project was designed with idea to allow a human player to play with AI.