shilangyu / TicTacToe

A simple tictactoe game with a trained AI

TicTacToe

Try me! Project where I created a trainer and a previewer of a TicTacToe AI. This project did not abuse the OP neural networks.

---

Preview

• npm start
• open browser in localhost:8080

Runs in the browser and draws the board on a canvas. The preview uses the previously trained json file as the "brain" of the AI. Try it out and see if you can beat it :)

Trainer

• npm run train -- <amount of games> <flag?>

  • --player AI will learn games where the player starts
  • --AI AI will learn games where the AI starts

The AI bases every move on the current tictactoe board. Each board can be represented in 8 same states: 4 rotations and 4 mirrors, which in result cuts the amount of possible combinations of a board by a factor of 8. Heres how the AI learns:

1. Parse the current board into string
2. Rotate and mirror the string (we have now 8 strings representing the same board)
3. Check if the json file contains any of the 8 strings
   • yes => get the x and y from the json file - make the x, y move
   • no => add a property to the json file containing all the possible moves for this board - make a random move
4. wait for the player move
5. check the current board state:
   • player won => forget about the previous move you've made
   • AI won => make the previous move your only one
6. goto 1