Core Concepts
Accelerating research into Unsupervised Environment Design (UED) with JaxUED library.
Abstract
Introduction
JaxUED provides fast, clear, and easily modifiable implementations of UED algorithms.
Leverages hardware acceleration for significant speedups.
Unsupervised Environment Design
UED involves generating environment distributions for robust policy learning.
Two-player game framework with a student policy and an adversarial level generator.
The JaxUED Library
Minimal dependency design inspired by CleanRL.
Introduces UnderspecifiedEnv interface for decoupling level distribution from environments.
Reference Implementations
Includes Domain Randomisation, Prioritized Level Replay, ACCEL, and PAIRED methods.
Results
Performance comparison with other libraries like DCD and minimax.
Surprising effectiveness of Domain Randomisation in UED methods.
Stats
JaxUED achieves on the order of 100× speedup compared to prior CPU-based implementations.
Quotes
"We aim to accelerate research into UED by making high-quality implementations available."
"DR is competitive with state-of-the-art UED methods."