Learning to Elect

Authors: Cem Anil, Xuchan Bao

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we investigate the effectiveness of PIN models on 1) mimicking classical voting rules (Section 4.1) and 2) discovering novel voting rules that maximize social welfare (Section 4.2). Table 1: Voting rule mimicking accuracy of learned voting rules.
Researcher Affiliation Academia Cem Anil University of Toronto Vector Institute anilcem@cs.toronto.edu Xuchan Bao University of Toronto Vector Institute jennybao@cs.toronto.edu
Pseudocode Yes Algorithm 1: Supervised Learning of Voting Rules
Open Source Code No The paper does not contain any explicit statements about the release of source code or a link to a public repository for the methodology described.
Open Datasets Yes We used synthetically generated elections to train the networks. We trained the networks on synthetic data to mimic the classical voting rules, and tested them on three real-world datasets: the Sushi dataset [Kamishima, 2003] (10 candidates), the Mechanical Turk Puzzle dataset [Mao et al., 2013] (4 candidates) and a subset of the Netflix Prize Data [Bennett et al., 2007] (3-4 candidates).
Dataset Splits No The paper describes training and testing phases but does not explicitly provide specific training/validation/test dataset splits (e.g., percentages or counts) or methodology for creating a distinct validation set.
Hardware Specification No The paper mentions the use of 'GPUs' in a footnote ('batches of elections can be parallelized very efficiently in PINs with GPUs.'), but does not provide specific hardware details such as exact GPU models, CPU models, or memory specifications used for the experiments.
Software Dependencies No The paper mentions software components and optimizers like PyTorch, Adam optimizer, Lookahead optimizer, and Layer Norm, but it does not provide specific version numbers for these software dependencies as used in their experiments.
Experiment Setup Yes We used the Lookahead optimizer [Zhang et al., 2019] to train the Deep Set models, and the Adam optimizer [Kingma and Ba, 2015] to train the other networks. We tuned the learning rate for each architecture. We used the cosine learning rate decay schedule with 160 steps of linear warmup period [Goyal et al., 2017]. We used a sample size of 64 elections per gradient step. We trained each PIN model for 320,000 gradient steps. We trained each MLP model for three times as long (960,000 gradient steps), as MLP models are observed to learn more slowly. Additional details for the training setup are included in the Supplementary Material.