Arthur Samuel
Arthur Lee Samuel (1901-1990) was an American electrical engineer and computer scientist who spent most of his career at IBM and is remembered as a founder of machine learning, for writing checkers programs that improved with experience and for the 1959 paper that described them. Born in Kansas and trained at MIT, he was an engineer first and a theorist second, which shaped everything about his approach: he wanted a working program that demonstrably got better, and he treated the game of checkers as a laboratory rather than as an end in itself.
His path to the work ran through hardware: after degrees from MIT, years at Bell Telephone Laboratories working on electronics, and a stint at the University of Illinois, where he took up the idea of a checkers program and a computer to run it, he joined IBM in 1949. At IBM in Poughkeepsie he worked on early machines, and when the IBM 701 became available in the early 1950s he began to program checkers on it. The 701's small memory forced economy: a position had to be encoded compactly, and the program had to play a move in a reasonable time. His checkers demonstrations became a well-known part of IBM's early computing story.
The paper that made his name, Some Studies in Machine Learning Using the Game of Checkers, appeared in the IBM Journal of Research and Development in 1959, and it is the standard source for the phrase machine learning in its modern sense. It described two learning mechanisms. Rote learning stored the backed-up value of each position the program had searched, so that when the position recurred the program could skip the work and effectively search deeper. Learning by generalization adjusted the coefficients of an evaluation function, a weighted combination of board features, by comparing the program's static evaluations with the values returned by lookahead search: when the two disagreed, the weights shifted to reduce the discrepancy. In effect, the program used its own deeper analysis as a teacher for its shallower judgment, an idea that foreshadowed temporal-difference learning in reinforcement learning, though Samuel did not use that name.
Self-play was the engine of the improvement: Samuel arranged for one copy of the program, with its current parameters, to play against a copy with altered ones, and the version that won was kept, so the program could gain experience without a human opponent. He was open about the limits, though. In his 1959 and 1967 papers he noted that the program stayed far from master level, that it needed a human-chosen set of features, and that it could not invent new features on its own, a weakness that later researchers continued to wrestle with. He also contributed to the development of search practices in game programming, including the use of cutoffs that resemble alpha-beta pruning. A later chapter of his career was at Stanford, where he moved in the 1960s and where, in later life, he wrote an introductory manual for Donald Knuth's TeX system.
Several misconceptions cling to Samuel's story: that he invented the phrase to mean what it means today, that his program was a champion, and that it learned unaided from nothing. The phrase machine learning appears in his title, and he is conventionally credited with popularizing, and probably coining, the term in its current sense, but historians are cautious about claiming absolute priority. The program was a competent player that beat a good amateur in a famous 1962 exhibition, not a champion. And its learning stood on human foundations: the rules, the feature list, the search procedure and the training scheme were all designed by Samuel. Today, his central idea, that a system can improve its own evaluation by comparing predictions against better-informed later ones, runs through value learning in modern reinforcement learning and self-play systems.
Read Arthur Samuel as the engineer who showed, on a primitive mainframe and with a children's board game, that a program's competence could come from its own experience rather than only from its author's foresight.