CONCEPT
Self-Play Learning
The method Arthur Samuel invented in 1959 in which a learning system generates its own training experience by playing against copies of itself—automatically calibrating difficulty to its own level and discovering strategies its author never specified.
Self-play is the method by which a learning system escapes the dependency on external training data or human opponents and manufactures its own curriculum from the rules of a problem.
Arthur Samuel invented it in 1959 out of hard practical necessity: strong human checkers opponents were scarce and slow, and a program that could only learn from games against masters would improve glacially. His solution was to let the program play against copies of itself—one version holding its evaluation fixed as a stable benchmark while the other adjusted its weights, then promoting the improved version and starting again. The machine bootstrapped itself up by its own results, with no human input beyond the rules of the game. The elegance is that the difficulty of the training automatically tracks the learner’s ability: too-weak opponents teach nothing, too-strong opponents punish without instructing, and a self-play system is always playing someone roughly its own strength. The line from Samuel to the present is