The board consists of 9 squares. Each of those is modeled by an input neuron. An empty square receives a value of 0. A square with our symbol on it receives a value of 1. A square with the opponent's symbol on it receives a valu of -1.

The agent makes a move by setting one of the nine output neurons to a value of at least 1. All other output neurons must be set to values less than 1.

If the output is not a valid move (fewer or more than one square is requested, or the requested square is already occupied), this is counted as a loss.

Winning the game awards the agent a score of +1. A draw awards the agent a score of +0.25. In addition, for any outcome besides a win, the agent is awarded 0.25-0.25/(1+nmoves) points.