I mentioned Menace the other day. This was a “Learning Machine” implemented entirely with Matchboxes. This was back in the 1960s. The idea is that a simple game like noughts and crosses (aka Tic-tac-toe) can be played by a crude form of learning computer.
You need enough matchboxes for every possible position in the game. That’s 304 matchboxes and you start but putting four coloured beads into each box for every possible turn. These represent the possible moves. To make a move, you pick the box corresponding to the position shake it and draw a bead out and play that move.
And so on until the game is finished. Then if Menace (as it was called- Machine Educable Noughts And Crosses Engine) won, you add three beads to each winning Matchbox, put one back in if its a draw and remove all of that colour if it lost. Gradually it becomes better at winning.
There’s a far detailed write up here. It’s an intriguing technique though possibly not the easiest to apply as it requires you to have a matchbox for every position in a game. Fine with noughts and crosses but impossible with chess. Instead of Matchboxes, it could be done way faster in software.
It needs games with determinate moves and not too many, though I suppose with RAM in modern PCs, a few million moves would be manageable. One possibility would be the L game. I’ve never heard of anyone doing that, and it would make for an interesting project. I’ll add it to my list of future projects. Now I wonder how many possible moves there are in a L- Game? It’s made more complex because you can optionally move one of the single pieces as well as the L-Shape which can be rotated and flipped and it’s a 4 x 4 board not 3 x 3 as in noughts and crosses. .