Can a machine learn from experience? Of course they do, we have artificial intelligence computer programs that learn from experience, from people, and from other computers. But all that had to start somewhere. In 1961, Donald Michie built a device called MENACE, which stands for Machine Educable Noughts And Crosses Engine (Noughts and Crosses is known as Tic-Tac-Toe in America). It was made out of a bunch of matchboxes and a supply of glass beads. In 2010, artist Julien Prévieux built a nice version of that same machine, called MENACE 2, with tiny drawers that resemble a library card catalog and a huge supply of colored beads. But what’s really mind-blowing is how it works. Many young engineers have recreated the project, but it’s new to me, and is a nuts-and-bolts lesson in how machines learn.
There are 304 little wooden drawers (or matchboxes in the original version created by Michie.) Each of them represents a unique board position that the player can encounter during a game. Each drawer is filled with coloured beads that represent a different move in that board state. The quantity of a colour indicated the “certainty” that playing the corresponding move would lead to a win.
Menace “learns” to win the game by playing repeatedly against the human player, honing its strategy until its opponent is only able to draw or lose against it. The trial and error learning process involves being “punished” for losing and “rewarded” for drawing or winning. This type of machine learning is called reinforcement learning.
To explain the process, we are led through a game of Tic-Tac-Toe and the consequences of winning or losing. Oh yeah, it’s slow and tedious, but it works, and eventually MENACE will defeat almost any player (although I wonder if there's been games played between two such devices). You can see how electronic computers can do this much quicker, but you’ll also see how the human brain is still much better at learning. -via Metafilter
(Image credit: Jousse Entreprise)