AI Glossary Term

What does “Self-Play” mean?

Definition

A training setup where a model improves by generating data through interactions or competitions with copies of itself.

Browse the full AI Glossary Explore all free AI guides

How Self-Play works

Example

AlphaGo Zero learned Go by playing millions of games against itself, starting with only the rules, and eventually surpassed earlier versions that had learned from human games.

Why it matters

Self-play creates unlimited training data in games and simulations. It works best where outcomes are clear and rules are fixed.

Common questions

Where is it used?

Board games, video games, and some research on reasoning and negotiation.

Why does it work?

The opponent improves as the learner does, so challenges stay at the right difficulty.

Limitations?

It is harder to apply where there is no clear winner or the rules of the world are fuzzy.

Scaling LawAn empirical relationship showing how performance improves with model size, data, or compute.Self-Supervised LearningLearning representations from unlabeled data by predicting masked or transformed parts.Synthetic DataArtificially generated data used to augment, simulate, or protect sensitive training data.BackpropagationThe core training algorithm that updates model weights by propagating prediction errors backward through the network.Continual LearningTraining approaches that let a model keep learning from new data without forgetting prior knowledge.Data AugmentationTechniques that create modified training examples to improve model generalization.

Learn more in our free guides

BYOL and Non-Contrastive Self-Supervision Pseudo-Labeling and Self-Training Self-Consistency Decoding Toolformer and Self-Taught Tool Use

Test yourself

BYOL and Non-Contrastive Self-Supervision Quiz Pseudo-Labeling and Self-Training Quiz Self-Consistency Decoding Quiz