Nonparametric Scoring Rules

Authors: Erik Zawadzki, Sebastien Lahaie

AAAI 2015 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results are provided that confirm rapid convergence and that the expected score correlates well with standard notions of divergence, both important considerations for ensuring that agents are incentivized to report accurate information. We conducted experiments to investigate the empirical properties of the sample-based kernel score.
Researcher Affiliation Collaboration Erik Zawadzki Carnegie Mellon University Pittsburgh, PA 15213 epz@cs.cmu.edu S ebastien Lahaie Microsoft Research New York, NY 10011 slahaie@microsoft.com
Pseudocode No The paper does not contain any pseudocode or clearly labeled algorithm blocks.
Open Source Code No The paper does not provide any concrete access to source code for the methodology described.
Open Datasets No The paper describes generating synthetic data for its experiments: "Our experiments require a way to generate a distribution P for the ground truth of an event and a distribution Q for the agent s beliefs. We generate P and Q independently from the same distribution of distributions. We used a mixture of between five and ten isotropic Laplace densities with bandwidth 0.05. Centers for the individual Laplace densities were located in [ 1, 1]D uniformly at random, and had weights drawn uniformly from [0, 1]." This is not a publicly available dataset with concrete access information.
Dataset Splits No The paper describes how samples and mixtures were generated and used for evaluation (e.g., "For each Pi, we ran M = 250 test instances. Each instance j M consisted of generating two m = 10, 000 sample sets Xi,j, X i,j P m i . For each instance the agent reported a prefix of k elements of Xi,j and was evaluated against the set X i,j."), but it does not specify traditional train/validation/test splits on a fixed dataset. Data is synthetically generated for each instance.
Hardware Specification Yes All experiments were coded in MATLAB, and run on a 3.40GHz i5-3570K with 8GB RAM.
Software Dependencies No The paper states, "All experiments were coded in MATLAB," but does not specify a version number for MATLAB or any other software libraries or dependencies with their versions.
Experiment Setup Yes Kernel bandwidth was 0.25, binning used 43 bins. The settings used in this experiment were found through an initial set of calibration experiments. In 1D roughly 45 bins seemed reasonable, whereas 62 bins was the best in 2D, and 43 bins was the best in three dimensions. We used 0.25 as our bandwidth in all three dimensions. The setting of m1 = 1500 was intended to represent a moderately sized report.