Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].

Near-Optimal Multi-Agent Learning for Safe Coverage Control

Authors: Manish Prajapat, Matteo Turchetta, Melanie Zeilinger, Andreas Krause

NeurIPS 2022 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We extensively evaluate our algorithms on synthetic and real problems, including a biodiversity monitoring task under safety constraints, where SAFEMAC outperforms competing methods. This section compares MACOPT and SAFEMAC to existing methods (or their extensions) on synthetic and real-world problems. We validate our theoretical claims and observe their superiority.
Researcher Affiliation Academia Manish Prajapat ETH Zurich EMAIL Matteo Turchetta ETH Zurich EMAIL Melanie N. Zeilinger ETH Zurich EMAIL Andreas Krause ETH Zurich EMAIL
Pseudocode Yes Algorithm 1 Greedy UCB (GREEDY), Algorithm 2 MACOPT, Algorithm 3 Safe Expansion (SE), Algorithm 4 SAFEMAC
Open Source Code Yes Joint supervision. Code available at https://github.com/manish-pra/SafeMaC
Open Datasets Yes The nest density is obtained by fitting a smooth rate function [23] over Gorilla nest counts [24]. As a proxy for bad weather, we use the cloud coverage data over the KGS from Open Weather [22].
Dataset Splits No The paper does not specify explicit training, validation, or test splits for any dataset. It describes experiments in simulated environments or using derived data, where 'samples' are collected iteratively during algorithm execution rather than from predefined splits of a static dataset.
Hardware Specification No The paper states 'Compute details are in Appendix G' in its checklist, but Appendix G is not provided in the given text, thus specific hardware details such as GPU/CPU models or memory amounts are not accessible.
Software Dependencies No The paper cites software like 'Botorch' [67], 'Gpytorch' [68], and 'Pytorch' [69] in its references, implying their use. However, the main text does not explicitly state specific version numbers for these or other software dependencies required to reproduce the experiments. Details are referred to Appendix G, which is not provided.
Experiment Setup Yes We perform our experiments with N = 3 agents in a 30 30 grid world where states are evenly spaced over [0, 3]2. Each agent s disk is defined as the region an agent can reach in r = 5 steps in the defined grid. We set βq = 3 and βρ = 3 for all t 1. In synthetic data, both the density ρ and the constrain q are sampled from a GP with zero mean and Matérn Kernel with ν = 2.5, scale σk = 1, and lengthscale l = 2. The observations are perturbed by i.i.d. Gaussian noise, N(0, 10 3). In obstacles, we use q (v) = 1/(1 + exp( 1.5dm(v))), to map the distance between [0, 3] and saturate the constraint value for large distances, and we set q(v) = q (v) 0.5 to avoid collisions.