self-play

appears in 2 paper titles

Definition

Training a model against copies of itself and learning from the resulting games. In its pure form nothing beyond the rules is required, and the opponent improves in lockstep with the learner, so the curriculum stays at the right difficulty automatically. Some systems instead bootstrap from human games and switch to self-play later, so the method does not by itself imply zero human data. The idea now extends past games, wherever a model can generate its own problems.

Explainers using this term

Self-Improving AI — Self-Play, Co-Evolution, and Generated CurriculaMastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm