About me
Skills
Experience
Interests
E-mail me!
corner
corner
corner
corner
border
border
border
border

Noughts and Crosses with a Neural Network

Well, two neural networks actually, both simple multi-layer perceptrons. Both of these have been trained using back propagation on 20 games of noughts and crosses (tic-tac-toe), one on all the possible combinations of board positions and moves that led to a win, and the other on all the combinations that led to a lose (drawn games weren't recorded). During any game, the current board position forms the input to both of these neural networks and allowed to propagate forward to the outputs. The score for every possible next move is a combination of the appropriate outputs from the two networks.

The networks have been deliberately only partially trained so that the computer plays a reasonable game but is not unbeatable.

Click here to play