The machine made of matchboxes that learns to play tic-tac-toe — with no electricity. (Donald Michie, 1961)
Each matchbox matches one situation in the game. Inside, there are colored marbles: one color = one square where MENACE can play. On the board, each square wears its color (red = square 1, orange = square 2…).
To play, MENACE shakes the box and draws a marble without looking. The color says which square to play. The more marbles a color has, the more chances it has to come out.
At the end of each game, we reopen only the boxes that were used and, in each one, we change the marble that was drawn (the color of the square that was played):
The marble we change is green here (square 4): it's the color MENACE played in this box. We touch neither the other marbles in the box, nor the ~300 other boxes.
Michie's original setting: at the start, 4 marbles per square at move 1, then 3, 2, 1 for the next moves. Lots of marbles = slow, careful learning (the frequent, subtle moves of the beginning); few marbles = fast learning (the rare, obvious moves of the end).
MENACE plays the ✖ and goes first. You play the ◯.
Make MENACE play super fast against a random opponent, and watch it stop losing.
One library per move. There are more and more possible situations, so more and more boxes: 1 → 12 → 108 → 183 (304 in all). Click a library, open a box, draw/add/remove a marble, then play to go down to the next move.
Keep an eye on the “first box”: 36 → 39 → 38 → 39. (+3 win, +1 tie, −1 loss, on the color that was played.)
Made to discover AI while having fun • Inspired by MENACE, Donald Michie (1961)