Brian Hayes in American Scientist:
Guided by an unseen hand, a grimy railroad tank car negotiates a series of switch points in the tracks, veering right, then right again, then left. Next comes a lime-green box car, which makes two lefts. I observe these events from the control tower of a railroad facility called a hump yard, where freight cars sort themselves into trains bound for various destinations. It is an eerie scene. The cars glide silently downhill through the maze of tracks, seeming to steer themselves, as if each car knows just where it wants to go. This is an illusion; a computer two floors below me is making all the decisions, setting the switches a moment before each car arrives. But I can’t shake the impression that the hump yard itself is a kind of computer—that the railroad cars are executing some secret algorithm.
It’s not such a far-fetched notion. In 1994 Adam Chalcraft and Michael Greene, who were then at the University of Cambridge, and later Maurice Margenstern of the University of Metz, designed railroad layouts that simulate the operation of a computer. The machine is programmed by setting switch points in a specific initial pattern; then a locomotive running over the tracks resets some of the switches as it passes; the result of the computation is read from the final configuration of the switches.