Tic-tac-toe, one network

You play X. The computer plays O, and it never loses.

What's new: in the original demo, a trained network picked the moves, and rules written by hand sat around it: “this square is taken”, “X has won”. Here those rules became part of the network. The whole player is one network.
What you'll notice: nothing different, and that's the point. It plays exactly like the original, move for move. The rules turned into numbers without changing anything.
Try: play a game. Every move goes through that one network.

Your moveYour turnThe network winsYou win?!Draw

No rule is written in the CSS: the network decides everything, including when the game is over.

How it works (for the curious)

The player, layer by layer

1,243 units and 5,434 nonzero weights. The trained layers come first; everything after the logits was written by hand. Each layer only reads the one before, so values needed later pass down through carry units, max(0, x) = x.

layer 1

layer 2

389

layer 3

515

layer 4

227

layer 5

layer 6

output

learnedhand-writtencarry (passes a value down a layer)

How a rule becomes weights

Rule	As ReLU
x ≥ t, for an integer x	`relu(x − t + 1) − relu(x − t)`
a AND b AND c	`relu(a + b + c − 2)`
a OR b OR …	`s − relu(s − 1)`, with s = a + b + …
a line of three X	`relu(x₀ + x₁ + x₂ − 2)`
square c is the best free one	`relu(1 − Σ_d beats(d, c))`, where d beats c if it is free and c isn't, or both are and d's score is higher
O moves only if you moved, didn't win, and the board isn't full	`relu(best_c + moved − x_won − full − 1)`

Checked in numpy on all 5,478 reachable boards: the network's move equals the original demo's integer model on all 2,097 positions where O is to move, and the judge net's winner, draw and free squares are right on every board. The browser test replays the original fixture games in Chromium, Firefox and WebKit. Built by handwired/scripts/rules.py and build_games.py.