Samuel's checkers program coins 'machine learning'

Arthur Samuel began writing a checkers program at IBM in 1952 because he wanted a problem simple enough to fit in an IBM 701 and hard enough to be interesting. By 1955 it was learning. The program scored positions with a weighted sum of features, piece advantage, mobility, control of the centre, and adjusted those weights by comparing its own predictions with what happened a few moves later. It played thousands of games against copies of itself, and it got better than Samuel, who was not a strong player, and eventually better than most people.

His 1959 paper in the IBM Journal describes two methods. Rote learning stored positions it had seen. Generalisation learning, the interesting one, changed the evaluation weights from experience. The paper is the first to use the phrase "machine learning" in print, and it defines it as the field still does: giving computers the ability to learn without being explicitly programmed.

The technique of learning from the difference between successive predictions is what Richard Sutton formalised as temporal-difference learning in 1988, and self-play against copies of itself is how AlphaGo Zero trained in 2017. Samuel's program was demonstrated on television in 1956; IBM's stock is said to have risen the next day.