Redes adversárias generativas: uma alternativa para modelagem de dados de entrada em projetos de simulação

Abstract

In general, stochastic simulation consists of input data and logic, the former being the basic source of uncertainty in a simulation model. For this reason, data modeling is an essential step in the development of stochastic simulation projects. Many advances have been observed in recent years in simulation software and in data collection tools. However, the methods for input data modeling have remained largely unchanged for over 30 years. In their daily lives, modelers face difficulties related to the choice of input data models, mainly due to the challenge of modeling non Independent and Identically Distributed Data (IID) data, which requires specific tools not offered by simulation software and their data modeling packages. For this reason, few studies consider elements of complexity such as heterogeneities, dependencies, and autocorrelations, underestimating the uncertainty of the stochastic system. Given the new developments in Artificial Intelligence, it is possible to seek synergies to solve this problem. The present study aims to evaluate the results of the application of Generative Adversarial Networks (GANs) for input data modeling. Such networks constitute one of the most recent architectures of artificial neural networks, being able to learn complex distributions and, therefore, generate synthetic samples with the same behavior as real data. Therefore, this thesis proposes a method for Input Data Modeling based on GANs (MDE-GANs) and implements it through the Python language. Considering a series of theoretical and real study objects, the results are evaluated in terms of representation quality of the input models and comparisons are made with traditional modeling methods. As a main conclusion, it was possible to identify that the application of MDE-GANs allows obtaining input data models with strong accuracy, surpassing the results of traditional methods in cases of non-IID data. Thus, the present thesis contributes by offering a new alternative for input data modeling, capable of overcoming some of the challenges faced by modelers.