Thanks to a recent episode of QI, I am finally able to explain machine learning to people. The show included a visit from Matthew Scroggs, a Research Fellow in the Department of Mathematics at University College London. Scroggs demonstrated a MENACE machine (a classic researcher backronym, in this case standing for the Matchbox Educable Noughts And Crosses Engine). The original MENACE machine was built by Donald Michie, an AI researcher who had worked alongside Alan Turing during the War.
MENACE is a pile of matchboxes with which you play tic-tac-toe (or noughts-and-crosses as we British call it). MENACE learns as it goes along which moves are “bad” moves, i.e., which moves lead to losing the game. It gradually removes “bad” moves from its portfolio of options, so with each game, it has more options for “good” moves, and fewer options for “bad” moves. This means it is more likely to pick good options, which means it’s more likely to win. So it appears to “learn”.
Here’s an older video of Matthew Scroggs demonstrating MENACE at the MATRIX Conference in Leeds (filmed by MathsWorldUK and the University of Leeds).