Grounding Aleatoric Uncertainty for Unsupervised Environment Design

Authors: Minqi Jiang, Michael Dennis, Jack Parker-Holder, Andrei Lupu, Heinrich Küttler, Edward Grefenstette, Tim Rocktäschel, Jakob Foerster

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We prove, and validate on challenging domains, that our approach preserves optimality under the ground-truth distribution, while promoting robustness across the full range of environment settings. Our experiments first focus on a discrete, stochastic binary choice task, with which we validate our theoretical conclusions by demonstrating that CICS can indeed lead to suboptimal policies.
Researcher Affiliation Collaboration Minqi Jiang UCL & Meta AI Michael Dennis UC Berkeley Jack Parker-Holder University of Oxford Andrei Lupu MILA & Meta AI Heinrich Küttler Inflection AI Edward Grefenstette UCL & Cohere Tim Rocktäschel UCL Jakob Foerster FLAIR, U of Oxford
Pseudocode Yes Algorithm 1: Sample-Matched PLR (SAMPLR)
Open Source Code Yes The code reproducing the experimental results is included in the supplemental material.
Open Datasets No The paper describes custom environments built on existing frameworks (Mini Hack, Car Racing Bezier) but does not provide concrete access information (e.g., link, DOI, or specific dataset citation) for a publicly available dataset used for training.
Dataset Splits No All agents are trained using PPO [42] with the best hyperparameters found via grid search using a set of validation levels. However, no specific details about the size or percentage of this validation set are provided.
Hardware Specification No The paper mentions "compute estimates in Appendix C" but does not explicitly provide hardware specifications (e.g., GPU/CPU models, memory details) in the main text provided.
Software Dependencies No The paper mentions software components like PPO, Mini Hack, Net Hack Learning Environment, and Car Racing Bezier environment, but does not provide specific version numbers for any of them.
Experiment Setup Yes We provide extended descriptions of both environments alongside the full details of our architecture and hyperparameter choices in Appendix C.