What it is
The formal branch of open-endedness: an environment generator and a learning agent set against each other so that levels stay hard but solvable. PAIRED defined UED with regret as the objective; POET coevolved environment and agent pairs; PLR showed curating random levels by learning potential works without controlling the generator; Robust PLR gave that a minimax-regret guarantee; ACCEL edits high-regret levels; the optimisation-framework paper replaced regret with learnability and proved convergence. Split out of Open-Endedness on 2026-09-10.
(agent-drafted from the sources’ summaries, not yet reviewed)
Key sources
- Emergent Complexity and Zero-shot Transfer via Unsupervised Environment Design — PAIRED; defines UED and introduces regret (antagonist − protagonist return) as the curriculum objective
- Paired Open-Ended Trailblazer (POET) — POET; environment-agent pairs coevolve with transfer, minimal criterion instead of regret; stepping stones are unpredictable
- Prioritized Level Replay — PLR; curate random levels by TD-error learning potential, needs no generator control; emergent easy-to-hard curriculum
- Replay-Guided Adversarial Environment Design — Robust PLR / DCD; gives PLR a minimax-regret guarantee and shows training on fewer (curated) levels transfers better
- Evolving Curricula with Regret-Based Environment Design — ACCEL; evolutionary edits of high-regret levels inside a Robust-PLR curator, the strongest practical UED method
- An Optimisation Framework for Unsupervised Environment Design — first convergence guarantee for UED via two-timescale GDA; generalised learnability replaces regret as the score
Related
- Open-Endedness
- Environment Generation
- Reinforcement Learning - RL
- UED Core Reading (personal notes, not in this repo)
To ingest
(none)