Basically, the game was a simpler version of tic tac toe, so one didn’t need to make 304 different match boxes in order to make a learning machine. Instead it was a project that probably only required 20-30 matchboxes, making it much simpler for an after-school group or kid on a rainy day to make and learn about.

If anyone can recall the simpler version of the game, I’d love to hear it.

It also teaches the basics of machine learning, the idea that “memory” might be implemented using beans, and even a rudimentary idea of an evolutionary algorithm (the original set with all the beans represents a large genetic pool; the fitness is the ability to win the game; and the system evolves to find the perfect solution).

Also, it’s like a mechanical computer. A difference engine. What’s not to like?

My scholarship is personal and self directed and I sometimes take a long time to discover what I’m looking for. I have imagined but failed to produce something quite like this for years…

why bother? because this method is extensible. If that doesn’t tickle your brain then take up gardening or something, because the promise(threat?) of machine learning speaks for itself to me. This example is of unusual explicative quality because many people can puzzle out the strategy of TTT, connecting to a ruleset which needs no teaching.

I’m still hoping to actually get a peek at one of these old paper computers I sometimes hear of as well.

It all seems anchored around the experience of reading Magister Ludi while I was out to sea once…

Joan Saberhagen

http://www.davincigames.it/giocarea_eng/10/profondita.htm

Anybody remember what book this was from?

http://www.mqasem.net/cool_stuff/ETicTacToe/TicTacToe.html

So this approach only works for games where a perfect strategy exists… and surely Go is not one of those? My mind boggles at how many matchboxes you’d need for Go.

Formally: Go is a zero-sum finite perfect
information game; all such games have an equilibrium point in pure strategies.

Skipping the game theory jargon: Go has a perfect strategy, and in principle it’s findable by MEGoE, the Matchbox Educable Go Engine. Of course, Go is very deep and would require very many matchboxes. So we have no idea what the winning strategy is. But there is one. Either Black can always win, or White can, or both players can guarantee a draw.

http://en.wikipedia.org/wiki/The_Adolescence_of_P-1

(Makes me want to build Lego logic gates…)

I’m guessing you’d need a shit ton of beans too – since each matchbox must contain enough beans for each subsequent path to be tried once.

The first matchbox would need as many beans as there are games of Go.

I learned about it from a Martin Gardner book.