To train the computer to play Tic Tac Toe we'll need to generate records of game play. The simplest way to start is for the computer to simulate both players making random (legal) moves. For every board position it records whether it eventually led to a win for X, O, or a tie. The trained model on each turn will then consider all possible moves and consider their relative scores. You can choose whether the trained computer player just picks the highest scoring move or uses the scores as probabilities of making that move.
Let's start by having the computer play games where both players just make random moves. You can choose how many games it will play in the settings panel that will appear at the top of this page.
How good is the computer at playing Tic Tac Toe now that you've trained it? One test is how well does it score the first move for X. The center and corners should score higher than the others.
Another way to evaluate the training is to have it play lots of games (you can decide how many in the Settings panel). You can have the trained model play against itself or against a player that makes random moves. If the player using the trained model always picks the highest scoring move then every game will be the same. Another strategy is to use the scores to determine the probability of making that move. High scoring moves are more likely than others but all moves can occur. If no moves have a positive score then all are considered likely to lead to a loss. In this case the least negative choice is selected.
You can save a trained model to the local file system to load at a later time.
Try different layer sizes for the model. Larger ones may lead to more accuracy at a cost of speed and memory usage. How does a narrow deep model compare with a wide shallower one? The learning rate is also worth exploring. One approach is to split the training up where each subsequent training session has a significantly smaller learning rate. Also set the number of iterations so that the learning is just levelling off so no need to waste time with more training. Experiment with different optimization methods and loss functions.
In addition to experimenting with different parameter settings, one can use games played by trained players to do further training. The default settings is that games created during the evaluation phase are added to the dataset of games available during training. You can also use the games for validation.
Challenge a friend or classmate (or all your classmates) to train the best Tic Tac Toe player. Run a thousand games for each contest to see whose player is stronger.