# Autocurriculum **Entity class:** Multi-agent learning mechanism **Domain:** Artificial intelligence / evolutionary dynamics **Maturity:** Developed ## Definition An **autocurriculum** is a sequence of learning challenges generated by the interaction of adaptive agents rather than prescribed in advance by a separate teacher. A successful policy changes what opponents must learn to counter it; their responses become the next challenge for the original learner. Cooperation can also generate new coordination problems as the group acquires capabilities it did not previously have. Leibo, Hughes, Lanctot and Graepel proposed this as a research hypothesis about innovation through social interaction, with increasingly complex tasks arising under particular conditions rather than automatically from any multi-agent system. ## A population as training environment An opponent can be both an agent and part of another agent's environment, which makes the training distribution endogenous. In [[wiki/Non-Transitive Games|non-transitive games]], training only against a single lineage can leave cyclic weaknesses; AlphaStar's StarCraft II league used main agents and specialized exploiters so that weaknesses discovered by one participant became training signals for others. The existence of a league is an engineered curriculum, while the more general autocurriculum names challenges arising from the learners' own interaction. Neither is evidence that every social encounter has been arranged by an external orchestrator. The route from here to [[wiki/Institutional Evolution|Institutional Evolution]] begins when participants can change the payoff structure or the rules of encounter, while [[wiki/Open-Endedness|Open-Endedness]] asks whether the sequence continues producing meaningful novelty rather than cycling through familiar responses. [[articles/Rock, Paper, Scissors|Rock, Paper, Scissors]] uses this distinction to connect the game to evolving social and machine-learning environments. ## Sources - [Leibo et al., *Autocurricula and the Emergence of Innovation from Social Interaction* (2019)](https://arxiv.org/abs/1903.00742) — research manifesto and definition. - [Vinyals et al., *Nature* (2019)](https://doi.org/10.1038/s41586-019-1724-z) — AlphaStar's multi-agent training league. - [Czarnecki et al., NeurIPS (2020)](https://proceedings.neurips.cc/paper/2020/hash/ca172e964907a97d5ebd876bfdd4adbd-Abstract.html) — transitive skill and cyclic strategic structure in games.