Arthur Lee Samuel (1901–1990) was an American electrical engineer and computer scientist, and a pioneer of artificial intelligence and machine learning. Working primarily at IBM and later at Stanford University, he developed checkers-playing programs that improved their performance through experience. These programs provided early experimental evidence that a computer could learn to perform a task better than its programmer. Samuel is also credited with coining the term “machine learning,” which appeared in the title of his influential 1959 research paper. (history.computer.org)
Education and engineering career
Samuel was born in Emporia, Kansas, and completed his undergraduate education at the College of Emporia. He earned a master’s degree in electrical engineering at the Massachusetts Institute of Technology in 1926, remained there as an instructor until 1928, and subsequently undertook graduate work in physics at Columbia University. He joined Bell Telephone Laboratories in 1928, where his research concerned electron tubes and related electronic devices. During World War II, he worked on switching devices that protected radar receivers from the powerful signals emitted by their transmitters. (history.computer.org)
In 1946, Samuel became a professor of electrical engineering at the University of Illinois and participated in its electronic-computer project. There he conceived a checkers program as a demonstration of what electronic computation could accomplish. He moved to IBM’s Poughkeepsie laboratory in 1949. His responsibilities included both existing vacuum-tube technology and research into switching transistors, linking his later work on learning programs to practical problems of computer construction. (history.computer.org)
Checkers as an experimental domain
Samuel used checkers, also called draughts, as a manageable but demanding research environment. Its rules and outcomes were clearly defined, while successful play required evaluating alternatives and anticipating an opponent’s responses. Games therefore offered a setting in which computer performance could be compared directly with human performance. Annotated records of expert games supplied another resource for experimentation. (legacy.cs.stanford.edu)
His program ran on the IBM 701 and subsequently on later IBM machines. Limited memory and processing time made exhaustive examination of all possible continuations impractical. The program instead searched selected portions of a game tree and used a heuristic evaluation function to estimate the merits of positions where search stopped. A minimax procedure propagated these estimates backward, choosing moves on the assumption that the opponent would also seek the most favorable continuation. (ibm.com)
Learning methods
Samuel’s 1959 paper, “Some Studies in Machine Learning Using the Game of Checkers,” distinguished two principal learning methods. Rote learning stored encountered board positions together with their computed scores. Reusing these records reduced repeated calculation and could effectively extend the program’s look-ahead. This was more than preserving a list of complete games: information about individual positions became available during subsequent searches. (gwern.net)
Learning by generalization altered the coefficients of the evaluation function. Board characteristics were represented by numerical features, and the program adjusted their relative contributions using evidence obtained during play and search. This allowed experience to affect the evaluation of positions that had not previously been stored. The distinction resembles the modern contrast between retaining specific cases and learning a reusable function approximation, although Samuel’s terminology and procedures preceded contemporary formulations. (gwern.net)
The experiments included self-play, in which versions of the program competed against one another. Samuel reported that a program could surpass its author’s playing ability after approximately eight to ten hours of machine-playing time under the conditions described in the paper. This result concerned improvement over the programmer, not superiority to the strongest human players. The system still depended on supplied rules, a direction for improvement, and human-selected candidate features. (ieeexplore.ieee.org)
Samuel also investigated learning from expert annotations. His program replayed material from Lee’s Guide to Checkers and adjusted its move-selection criteria toward choices judged good by experts. In modern terms, these annotated games supplied training data for a process resembling supervised learning, alongside learning through play. (legacy.cs.stanford.edu)
Results and subsequent development
In 1962, Samuel’s program defeated Robert Nealey, a human checkers player whose encounter with the machine became a prominent demonstration of learning-based computation. Its record against human opponents was nevertheless mixed; one publicized victory did not establish consistent master-level performance. (ibm.com)
Samuel’s 1967 follow-up paper described improved book learning, a signature-table technique, and more extensive use of alpha–beta pruning and other search restrictions. These changes permitted deeper look-ahead and improved play. The paper explicitly stated that the program remained unable to outplay checkers masters. It also identified limitations in the earlier linear evaluation method, including dependence on manually selected features and inadequate treatment of interactions between them. (teaching.bb-ai.net)
Samuel’s experiments became an early foundation for reinforcement learning, particularly temporal-difference learning. Their historical importance lies partly in updating position evaluations from other estimates generated during play and search, rather than relying exclusively on completed-game outcomes. Later accounts of temporal-difference methods identify his work among their early roots. (incompleteideas.net)
Stanford years and recognition
Samuel retired from IBM in 1966 and joined Stanford University as a lecturer and research associate; he became a research professor in 1974. He continued checkers research, worked on speech recognition, and remained active in teaching until 1982. His later programming contributions included the SAIL operating system, multiprocessor software, and the TeX typesetting system. He also wrote explanatory manuals, including First Grade TeX, reflecting a sustained interest in clear software documentation. (legacy.cs.stanford.edu)
In 1987, Samuel received the IEEE Computer Society’s Computer Pioneer Award. He died at Stanford Hospital on July 29, 1990. (history.computer.org)