Machine Learning
Checkers as a laboratory for intelligence. This is how Arthur Samuel conceived his approach. This pioneer, active from the 1940s through the 1960s, transformed computing by infusing it with the ability to learn from experience, a capacity until then reserved for humans.
The choice of checkers was by no means random. In the landscape of artificial intelligence, games played a role comparable to that of Drosophila for geneticists—simple organisms, easy to manipulate, perfect for experimentation. Checkers, less complex than chess, offered an ideal testing ground where Samuel could observe learning mechanisms without getting bogged down in overly sophisticated rules.
In 1952, the first program came to life on the IBM 701. Three years later, the version equipped with autonomous learning captivated the public during a televised demonstration. Without reaching the level of champions, this program innovated through two methods: rote learning and learning by generalization.
The first technique, rote learning, resembled our own memory more than one might imagine. Samuel programmed his computer to remember each configuration encountered during its games. Concretely, when the computer analyzed a position, it calculated its value using an algorithm called minimax that simulates the best possible moves for each player. The machine then stored this position with its evaluation. In subsequent games, faced with a situation already seen, no need for lengthy calculations—the answer sprang instantly from its memory. Samuel integrated an ingenious mechanism he called “depreciation.” To understand this, let us visualize a tree where each branch represents a possible move. The deeper we go into this tree, the further we advance into the hypothetical unfolding of the game. Samuel arranged for positions distant in this tree to be worth slightly less than those close to the present moment, even at equal game value. This subtlety encouraged the program to favor short paths to victory. Without this trick, the computer could have become mired in endless strategies, preferring to win in 50 moves rather than 10, simply because both scenarios mathematically led to victory.
His second method, learning by generalization, foreshadowed modern temporal difference learning techniques. Samuel had his program play against itself thousands of times. At each move, through a brilliant and intuitive idea, the computer subtly modified its way of evaluating the board. If a game position leads to other advantageous positions, then this initial position must be favorable. Imagine a move that seems mundane but systematically leads to winning configurations three moves later—the program gradually learns to value this apparently unremarkable move. The computer did not wait for the end of the game to adjust its evaluations; it learned continuously, after each movement. Without being explicitly told “this is good” or “that is bad,” the program discovered effective strategies on its own. Its compass remained nonetheless material advantage—the number of pieces owned relative to the opponent—with a marked preference for kings, those crowned pieces with superior movement capabilities. This approach, remarkably modern, constitutes one of the first manifestations of autonomous learning in 20th-century computing.
Beyond the framework of theoretical research, Samuel’s work on non-numerical computation shaped the architecture of early IBM computers. The logical instructions he proposed became an industry standard, their utility proving essential for all non-mathematical processing. Before revolutionizing artificial intelligence, he had distinguished himself as an electrical engineer. A graduate of Emporia College in 1923, then holder of a master’s degree from MIT in 1926, he taught there briefly before joining Bell Labs. His research on electron tubes, particularly on space charge between parallel electrodes, marked the era. During the war, his work on devices protecting radar receivers proved crucial.
In 1946, having become a professor at the University of Illinois, he participated in the creation of one of the first electronic computers. It was there that he dreamed of a checkers program capable of defeating a world champion. This project remained unfinished until his arrival at IBM in 1949, where he worked on the 701, the firm’s first stored-program computer. Samuel considerably improved its Williams tube memory, quadrupling its storage capacity and stabilizing its operation. The first public demonstration of his program so impressed Thomas J. Watson Sr., founder of IBM, that he predicted—correctly—a 15-point rise in the stock price.
Samuel contributed to IBM’s international expansion, shaping the direction of European laboratories, notably in Vienna (computing) and Zurich (physics). His discretion probably explains why the importance of his work was not fully recognized until after his departure from IBM in 1966.
At Stanford, where he became a research professor, he continued his work on checkers until other programs surpassed his in the 1970s. He became interested in speech recognition and supervised numerous doctoral students. His talents extended to technical writing. He excelled at understanding confusing documentation and producing clear manuals.
Samuel continued programming to a remarkable age. His last contribution, at 85, was to adapt multi-font printing programs for Stanford’s computer science department. Only Parkinson’s disease brought this exceptional activity to an end. His last login was on February 2, 1990, which probably makes him the oldest active programmer of his time.