PreferenceNet: Encoding Human Preferences in Auction Design with Deep Learning
Authors: Neehar Peri, Michael Curry, Samuel Dooley, John Dickerson
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct a number of experiments using synthetic data (as is typical for neural network based auction mechanisms [13]) and evaluate our method on different auction settings and fairness constraints. We also conduct two surveys to further study human preferences. We validate our approach through human subject research and show that we are able to effectively capture real human preferences. |
| Researcher Affiliation | Academia | Neehar Peri*, Michael J. Curry*, Samuel Dooley, John P. Dickerson Center for Machine Learning, University of Maryland peri@umiacs.umd.edu, {curry, sdooley1, john}@cs.umd.edu |
| Pseudocode | No | The paper describes the training algorithm in narrative text but does not include structured pseudocode or an algorithm block. |
| Open Source Code | Yes | Our code is available on Git Hub. We refer readers to Git Hub for our implementation. |
| Open Datasets | Yes | Given the lack of publicly available auction data, we generate synthetic bids as in [13, 25, 37]. All survey results are anonymized to protect participant privacy. We include these results in the supplemental material. |
| Dataset Splits | Yes | Rather, we evaluate each checkpoint on a validation set and maximize the following criteria in Eq. 2: α PCA + β p(b) / max(p(b)) + γ (1 − rgt(b) / max(rgt(b))) s.t. α + β + γ = 1. In our experiments we set α = 0.45, β = 0.1, γ = 0.45. |
| Hardware Specification | Yes | We run all our experiments on an NVIDIA Titan X (Pascal) GPU. |
| Software Dependencies | No | The paper mentions using the Adam optimizer but does not specify its version or any other software dependencies with version numbers. |
| Experiment Setup | Yes | For each configuration of n agents and m items, we train Regret Net for a maximum of 200 epoch using 160,000 training samples. We also train the MLP with 80,000 initial training samples and iteratively retrain the MLP with 5,000 additional samples from the partially trained Regret Net every 5 epoch. For all networks, we use the Adam optimizer and 100 hidden nodes per layer. We incremented ρr every 2500 iterations and λr every 25 iterations. In our experiments we set α = 0.45, β = 0.1, γ = 0.45. |