AI Hub
All terms

Self-Play

Improving by training against copies of oneself.

In self-play, an agent learns by repeatedly competing against versions of itself, generating its own ever-harder curriculum. It is how AlphaZero reached superhuman play in games with no human data.

Related papers