Probably Approximately Efficient Combinatorial Auctions via Machine Learning
Authors: Gianluca Brero, Benjamin Lubin, Sven Seuken
AAAI 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To evaluate the effectiveness of our proposed mechanisms, we have conducted a set of experiments to empirically identify the smallest PAE bound εδ for different distributions over value functions V, and for several ML-based mechanisms under different choices of inputs A, D, and q. In Tables 1 and 2 we present the results for δ = 0.1; we can therefore interpret the numbers provided as saying that 90% of the auction instances were at least this efficient. |
| Researcher Affiliation | Academia | Gianluca Brero Department of Informatics University of Zurich brero@ifi.uzh.ch Benjamin Lubin Information Systems Department Boston University School of Management blubin@bu.edu Sven Seuken Department of Informatics University of Zurich seuken@ifi.uzh.ch |
| Pseudocode | Yes | Algorithm 1: ML-Based Mechanism: One-Shot Version |
| Open Source Code | No | The paper does not include any explicit statement about releasing source code for the described methodology, nor does it provide a link to a code repository. |
| Open Datasets | Yes | We use two standard value models from the literature... The Global Synergy Value Model (GSVM) (Goeree and Holt 2008)... By contrast, the more complex Local Synergy Value Model (LSVM) (Scheffel, Ziegler, and Bichler 2012)... |
| Dataset Splits | No | The paper states parameters were 'tuned... from a hold-out set of 100 valuation profiles' and discusses 'initial training phase' with 'q samples'. However, it does not provide specific train/validation/test dataset splits (percentages or counts) for the main experimental data. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., CPU, GPU models, or memory) used for running the experiments. |
| Software Dependencies | No | The paper mentions using 'standard IP solvers (such as CPLEX)' but does not provide specific version numbers for CPLEX or any other software dependencies. |
| Experiment Setup | Yes | In our experiments we set c = 100 and ζ = 0.1 such that the minimization of the interpolation loss is prioritized in the objective of (10). We tuned these parameters from a hold-out set of 100 valuation profiles from the respective setting under consideration. |