Low-Variance Black-Box Gradient Estimates for the Plackett-Luce Distribution

Authors: Artyom Gadetsky, Kirill Struminsky, Christopher Robinson, Novi Quadrianto, Dmitry Vetrov10126-10135

AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate the effectiveness of the proposed method with a simple toy task similar to Tucker et al. (2017) and then continue to the more challenging task of optimization over topological orderings for solving causal structure learning problems.
Researcher Affiliation Academia 1National Research University Higher School of Economics 2Predictive Analytics Lab (PAL), University of Sussex
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks. It presents mathematical derivations but not in an algorithm format.
Open Source Code Yes Our Py Torch (Paszke et al. 2017) implementation of the gradient estimators is available at https://github.com/agadetsky/pytorch-pl-variancereduction .
Open Datasets No The paper mentions "We simulated graphs from two well-known random graph models" and "data sampled from the standard ALARM network", but does not provide concrete access information (link, DOI, specific repository, or formal citation to a dataset instance) for publicly available or open datasets.
Dataset Splits No The paper mentions "additionally generated validation samples Xval" and averages results across "5 random seeds", but does not provide specific dataset split information such as exact percentages, sample counts, citations to predefined splits, or detailed splitting methodology.
Hardware Specification No The paper mentions "NVIDIA for GPU donations" and "NRU HSE for providing computational resources", but does not provide specific hardware details such as exact GPU/CPU models, processor types, or memory amounts.
Software Dependencies No The paper mentions using "Py Torch (Paszke et al. 2017)", but does not provide specific version numbers for PyTorch or any other ancillary software dependencies.
Experiment Setup Yes For the PL-REBAR estimator we take cφ(z) = ηf(σ(z, τ)) where σ(z, τ) is the continuous relaxation of permutations described by Grover et al. (2019). For the PL-RELAX estimator we take cφ(z) = f(σ(z, τ)) + ρφ(z) where ρφ(z) is a simple neural network with two linear layers and Re LU activation between them. ... Regularization coefficient λ is set to 0.5 for all methods.