Arthur Samuel (1901-1990) was an American computer scientist at IBM who developed one of the first self-learning game-playing programs, a checkers-playing system, and coined the term "machine learning" in a landmark 1959 paper.
The checkers program
Samuel began developing his checkers program in the 1950s while at IBM, running it on the company's early computers, including the IBM 701. Rather than hand-coding a fixed strategy, Samuel designed the program to improve through experience: it played games against itself and against human opponents, adjusted the weights of its position-evaluation function based on outcomes, and used a rudimentary form of what would later be formalized as Reinforcement learning. A 1962 televised demonstration in which the program defeated a strong human checkers player drew wide public attention to the idea that a machine could improve its own performance without being explicitly reprogrammed.
Coining "machine learning"
In his 1959 paper "Some Studies in Machine Learning Using the Game of Checkers," Samuel defined the field as giving "computers the ability to learn without being explicitly programmed," a phrase that became the standard, if informal, definition of machine learning for decades afterward. The paper also introduced ideas resembling temporal-difference learning, in which the program adjusted its evaluations by comparing predictions at successive time steps, a technique later formalized and named by Richard Sutton and central to modern reinforcement learning.
Later career and influence
Samuel later worked at Stanford University, where he contributed to early time-sharing systems alongside John McCarthy and continued refining his checkers program into the 1970s. Though narrow by the standards of later game-playing systems, his checkers program anticipated the self-play paradigm that would resurface decades later in systems such as AlphaGo and, especially, AlphaZero, which learned Go, chess, and shogi from self-play alone.
Legacy
Samuel's checkers program is widely cited as the first computer program to learn to play a game well enough to challenge serious human players, and his coinage of "machine learning" gave the field a name that has endured even as its methods evolved from simple weighted evaluation functions to the deep neural networks and large language models of the twenty-first century. His demonstration that a machine could improve through experience rather than explicit instruction is often cited as an early proof of concept for the broader project of Artificial intelligence.