Enabling Adaptive Agent Training in Open-Ended Simulators by Targeting Diversity

Authors: Robby Costales, Stefanos Nikolaidis

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our empirical results showcase DIVA s unique ability to overcome complex parameterizations and successfully train adaptive agent behavior, far outperforming competitive baselines from prior literature. (from abstract) and 5 Empirical results (section title).
Researcher Affiliation Academia Robby Costales Stefanos Nikolaidis Department of Computer Science University of Southern California Correspondence to rscostal@usc.edu.
Pseudocode Yes Appendix A Algorithmic details. Algorithm 1 DIVA, Algorithm 2 DIVA (detailed), Algorithm 3 QD update.
Open Source Code Yes Our code is available at https://github.com/robbycostales/diva.
Open Datasets No The paper utilizes modified versions of environments like GRIDNAV, ALCHEMY [18], and RACING [17]. While these environments are referenced with citations to papers describing them, the paper does not provide direct links, DOIs, or specific repository names for publicly available datasets used in the experiments. It describes them as domains for their experiments rather than external, accessible datasets.
Dataset Splits No The paper describes training processes for agents and evaluation over environment distributions but does not specify explicit train/validation/test dataset splits (e.g., percentages or sample counts) for the data used in the experiments. It details meta-training and evaluation, but not data partitioning.
Hardware Specification Yes All results were produced on a handful of Titan X or Xp GPUs.
Software Dependencies No The paper mentions specific libraries and codebases used, such as 'pyribs', 'Vari BAD', 'PLR', and 'ACCEL', but does not specify their version numbers.
Experiment Setup Yes Table 4: DIVA hyperparameter settings. Table 5: Vari BAD hyperparameter settings.