How did a checkers program teach itself to beat a checkers master?

Arthur Samuel's checkers program taught itself by playing, including thousands of games against itself, until in 1962 it beat Robert Nealey, a self-described checkers master. Samuel, an IBM engineer, started the project in 1949. In 1959 he coined a name for what it did: machine learning. It means a program that learns from data and generalizes, instead of only following the steps it was given.

The program looked ahead through a tree of possible board positions. It scored each one with a formula weighing pieces, kings and how close pieces were to being crowned. Then it picked the move that was best, assuming the opponent also played its best. The learning came on top of that. It remembered positions it had already seen and whether they led to a win or a loss. Later versions also tuned the formula using professional games, and thousands of games against themselves.

The famous win needs a footnote. It was one game, and the program's later record against people was mixed. It reached a respectable amateur level, the first program to play any board game that well. But it was far from mastering checkers. A single win was enough to give the public the impression that it was very good.

Its core idea now runs through everyday tools. An email filter, for example, learns from labelled examples which folder each new message belongs in. That power comes with a classic trap: overfitting. A model that memorizes its practice data, noise and all, can ace every example it has seen and then fail on anything new, like a student who memorized one practice test.