So now Philip and I are exploiting our newfound knowledge about neural nets and evolutionary algorithms to create a neural net that plays tic-tac-toe.

We're both interested in game-playing AI, and this seemed like a nice, reasonably simple place to start.

We're currently working on a multi-layer network with nine input nodes and nine output nodes, corresponding to the nine spaces on the board:

3 A3 B3 C3

2 A2 B2 C2

1 A1 B1 C1

A B C


The input is the current board state, where 1 is the value for a friendly unit, 0 is a blank space, and -1 is an enemy unit. So an example input might be:

1 0 1

0 -1 -1

0 0 0

or: [1 0 1 0 -1 -1 0 0 0], the way the network looks at it.

The output will either be all zeros (pass), or all zeros and an output of 1 for the move.

So for example, an output of [0 0 0 0 1 0 0 0 0] would be a move in the center square of the tic-tac-toe board.

First we plan to evolve nets that understand the rules of the game, then we plan on evolving them to learn strategy.

I'll let you know how it goes.