Cayley Maze: Universal Open-Ended Reinforcement Learning Environment
Keywords: Reinforcement Learning, Unsupervised Environment Design, Open-Ended Learning
Abstract: Parametrizable environments with variable complexity are crucial for advancing fields such as Unsupervised Environment Design (UED), Open-Ended Learning, Curriculum Learning, and Meta Reinforcement Learning. However, the selection of environments in evaluation procedures, along with their complexities, is often either neglected or lacks formal justification. We propose the formal definition of complexity for Markov Decision Processes using Deterministic Finite Automata and Group Theory machinery. We introduce Cayley Maze, a novel open-ended reinforcement learning environment that naturally generalizes problems like solving Rubik's Cube, sorting, and integer factorization. Cayley Maze is universal: every finite deterministic sparse MDP is an MDP of a certain instance of Cayley Maze. We demonstrate how Cayley Maze enables control over complexity, simplification, and combination of its instances.
Primary Area: reinforcement learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 13937
Loading