Método para la síntesis de paisajes de optimización bidimensionales basado en redes adversarias generativas

Abstract

The implementation of an artificial neural network, specifically a Deep Convolutional Adversarial Adversarial Network (DCGAN), was carried out in order to generate two-dimensional optimization landscapes. The number of landscapes and their diversity were increased. We started with the creation of a database based on the literature, to later produce other databases with specific characteristics. Eight databases were created, the first four of which were divided into: Original, No noise, With a single minimum and With several minimums. The 4 subsequent databases were created from the application of a Data Augmentation to the 4 previously mentioned databases. Then, the algorithm was run several times until the appropriate parameters for the operation of the network were found, this was done for each of the databases. Once the parameters were chosen for each of the databases, the algorithm was run again several times to perform an analysis of the results and also obtained several optimization landscapes with different characteristics, with similarity to the original database and at the same time, with the presence of different information. From all the experimentation it is validated that it is possible to generate new two-dimensional optimization landscapes from a generative adversarial network, leaving as future work to find the functions or mathematical expressions that will produce such landscapes.