2284. Utility-Driven Hyperparameter Optimization for Tabular Data GANs via Ranking-and-Selection
Invited abstract in session WB-6: Predictive Analytics: Forecast Combination & Hyperparameter Optimization, stream Analytics, Data Science, and Forecasting.
Wednesday, 10:45-12:15Room: H9
Authors (first author is the speaker)
| 1. | Mick Molitor
|
| Information Systems Management, FU Berlin |
Abstract
A plethora of applications can benefit from the use of Generative Adversarial Networks (GANs) to generate samples, from data augmentation to solving complex optimization problems. Yet despite their versatility, GAN training poses a tough, often underreported, optimization challenge. Tabular GANs are notoriously brittle: small changes in weight initialization can significantly affect training and output quality. Traditional hyperparameter optimization (HPO) methods—such as grid search, random search, and Bayesian optimization—typically assume consistent outcomes across runs, often neglecting the substantial role of stochasticity in GAN performance compared to other neural architectures.
To better understand the intricacies of GAN training, HPO is framed as a stochastic ranking-and-selection (R&S) problem, and the Kim-Nelson indifference-zone procedure—well-established in discrete-event simulation—is adapted to GANs. Each hyperparameter vector is treated as an "alternative", with its mean performance estimated through repeated training runs using independent random seeds. The R&S method allocates additional replications only to candidates that remain competitive and halts once the probability of correct selection exceeds a user-defined threshold, providing finite-sample guarantees not available in standard HPO.
This contribution offers a meta-level perspective on hyperparameter optimization, emphasizing not only which configurations yield the highest utility metrics but also how those metrics relate to the robustness of model performance under stochastic variation.
Keywords
- Machine Learning
- Stochastic Models
- Predictive Analytics
Status: accepted
Back to the list of papers