Details

Autor(en) / Beteiligte

• Barra Montevechi, Jose Arnaldo
• Campos, Afonso Teberga
• Gabriel, Gustavo Teodoro
• dos Santos, Carlos Henrique

Titel

Input Data Modeling: An Approach Using Generative Adversarial Networks

Ist Teil von

• 2021 Winter Simulation Conference (WSC), 2021, p.1-12

Ort / Verlag

IEEE

Erscheinungsjahr

2021

Quelle

IEEE/IET Electronic Library

Beschreibungen/Notizen

• Input data modeling varies according to the modeler's objectives and may be a simple or complex task. Despite great advances in data collection techniques, the input data analysis remains a challenge, especially when the input data is complex and cannot be modeled by standard solutions offered by commercial simulation software. Therefore, this paper focuses on how Generative Adversarial Networks (GANs) may support input data modeling, especially when traditional approaches are insufficient or inefficient. We evaluate the adoption of GANs for modeling correlated data as well as independent and identically distributed data. As results, GAN input models were able to generate highly accurate synthetic samples (average accuracies> 97.0%). For univariate distributions, we found no significant difference between standard and GAN input models performances. On the other hand, for correlated data, GAN input models outperformed standard ones. The most relevant accuracy gain was observed for the bivariate normal.