Optimal Auctions through Deep Learning
Authors: Paul Duetting, Zhe Feng, Harikrishna Narasimhan, David Parkes, Sai Srivatsa Ravindranath
ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We prove generalization bounds and present extensive experiments, recovering essentially all known analytical solutions for multi-item settings, and obtaining novel mechanisms for settings in which the optimal mechanism is unknown. We present the complete set of experiments in the appendix and include a representative subset of the results here. We implemented our framework using the TensorFlow deep learning library. We used the Glorot uniform initialization (Glorot & Bengio, 2010) for all networks and the tanh activation function at the hidden nodes. For all the experiments, we used a sample of 640,000 valuation profiles for training and a sample of 10,000 profiles for testing. The augmented Lagrangian solver was run for a maximum of 80 epochs with a minibatch size of 128. |
| Researcher Affiliation | Academia | 1London School of Economics 2Harvard University. Correspondence to: Zhe Feng <zhe_feng@g.harvard.edu>. |
| Pseudocode | Yes | The solver is described in Algorithm 1. We divide the training sample S into mini-batches of size B, and perform several passes over the training samples (with random shuffling of the data after each pass). We denote the minibatch received at iteration t by St = {u(1), . . . , u(B)}. The update (a) on model parameters involves an unconstrained optimization of Cρ over w and is performed using a gradient-based optimizer. |
| Open Source Code | Yes | We implemented our framework using the Tensor Flow deep learning library.6 6https://github.com/saisrivatsan/deep-opt-auctions |
| Open Datasets | No | The paper describes generating valuation profiles by drawing values from uniform distributions (e.g., U[0,1], U[2,3]), rather than using a specific, named public dataset with concrete access information. |
| Dataset Splits | Yes | For all the experiments, we used a sample of 640,000 valuation profiles for training and a sample of 10,000 profiles for testing. For setting (VI), we show in Figure 5(a) the revenue and regret of the learned auction on a validation sample of 10000 profiles, obtained with different architectures. |
| Hardware Specification | No | Our experiments were run on a compute cluster with NVDIA GPU cores. This statement is too general and does not provide specific hardware details like GPU model numbers or CPU specifications. |
| Software Dependencies | No | We implemented our framework using the Tensor Flow deep learning library. We used the Adam optimizer (Kingma & Ba, 2014). The LP is solved with the commercial solver Gurobi. The paper mentions software tools like TensorFlow, Adam optimizer, and Gurobi, but does not provide specific version numbers for these software components. |
| Experiment Setup | Yes | We used the Glorot uniform initialization (Glorot & Bengio, 2010) for all networks and the tanh activation function at the hidden nodes. For all the experiments, we used a sample of 640,000 valuation profiles for training and a sample of 10,000 profiles for testing. The augmented Lagrangian solver was run for a maximum of 80 epochs with a minibatch size of 128. The value of ρ in augmented Lagrangian was set to 1.0 and incremented every 2 epochs. An update on wt was performed for every minibatch using the Adam optimizer with learning rate 0.001. For each update on wt, we ran R = 25 misreport updates steps with learning rate 0.1. |