A checkers program that learns reaches American television

What happened

Samuel began the program at IBM in the early 1950s and kept refining it for years afterward. Two ideas did most of the work. It searched ahead through possible moves and scored the resulting positions with a weighted formula, and it adjusted those weights according to the outcomes of games it played against a copy of itself. It also stored positions it had already evaluated so it wasted less effort on familiar situations. Checkers suited the purpose because the rules are simple while the tree of possibilities is far too large to examine completely.

The televised demonstration was corporate publicity, and effective publicity, since a machine winning a board game photographs better than a payroll calculation. Samuel published his methods later in the decade in a paper still cited as founding work in machine learning. The other episode people remember, a 1962 match in which the program beat a strong human player, has been retold with steadily inflating credentials for the opponent. Contemporaneous accounts describe an accomplished player rather than a national champion, and Samuel himself was cautious about how much the result proved.

Checkers kept its place as a measuring stick for decades. A program from the University of Alberta won a world championship match in the 1990s, and by 2007 the same group had shown that perfect play by both sides ends in a draw, which closed the game as a research problem. Samuel's contribution is best described not as the strongest player ever built but as the first convincing demonstration that a program could improve without a person telling it how.

The world at the time

Commercial computers in the mid-1950s were rented by large institutions and used for accounting, engineering and defense work. Artificial intelligence had barely been named, since the summer workshop that gave the field its label met in 1956. Games offered researchers a rare problem with clear rules, an unambiguous result and a difficulty that no amount of raw arithmetic could brush aside.

What it changed

Samuel's paper became a standard citation, and his approach of learning an evaluation function from results rather than instruction fed into decades of research. IBM kept using games as public proof of its capability, a line that runs to its chess and quiz show machines. For a general audience, the notion of a computer that learns entered circulation here.