Textbook of AI

1901–1990, Computer scientist at IBM

Also known as: Arthur Lee Samuel

Arthur Lee Samuel was an American computer scientist at IBM whose checkers-playing program, developed from 1952 onwards on the IBM 701 and 704, was the first computer program to demonstrate genuine machine learning in a non-trivial domain. The 1959 paper Some Studies in Machine Learning Using the Game of Checkers introduced the phrase machine learning and demonstrated two learning techniques that remain central: rote learning of board-position evaluations and generalisation by adjusting the coefficients of a polynomial evaluation function via a method that, in retrospect, is recognisably temporal-difference learning decades before Sutton's formalisation.

By 1962 Samuel's program defeated Robert Nealey, a Connecticut state checkers champion, in a single game, a result widely (and somewhat misleadingly) reported as "computer beats champion". The program could not in fact reliably defeat strong human players until much later, but the demonstration of learned competence was unprecedented.

Samuel's other contributions include the design of IBM's early symbolic assembler (with Nathaniel Rochester) and a long career on the IBM 701 hardware team. After retirement from IBM in 1966 he moved to Stanford as a professor and continued to refine the checkers program into the 1970s. His work was a direct ancestor of TD-Gammon (1992), of AlphaGo (2016), and of every reinforcement- learning agent that learns from self-play.