Redes neurais artificiais: uma aplicação em petrofísica e estudo dos efeitos de estímulos persistentes

• Ano de defesa: 2007
• Autor(a) principal: Vieira, Vinícius Manzoni
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Alagoas; Brasil; Programa de Pós-Graduação em Física; UFAL
• Palavras-chave em Português: Redes nuerais (Computação); Petrofísica; Hopfield, Modelo de; Estímulos persistentes; Nueral networks (Computing); Petrophysics; Hopfield, Model of; Persistent stimuli; CNPQ::CIENCIAS EXATAS E DA TERRA::FISICA
• Link de acesso: http://www.repositorio.ufal.br/handle/riufal/4763

Artificial Neural Networks are mathematical and computational devices which try to simulate some behaviors of biological neural networks. In the beginning of this work, we present a review about neural networks, showing an application of a Multi-Layered Perceptron for the estimative of petrophysical parameters. After that, we made an analytical and numerical study of the behavior of a neural network based on the Hopfield model on which we introduced in the system’s dynamics a parameter that acts as a field to simulate the effect of a persistent stimulus that privileges a stored pattern in the network. For the Hopfield’s model in the presence of this field, we present a review of the field’s effect in the model with dilution and asymmetry on the synaptic connections. After that, using a mean Field approximation, we have got a set of equations for the order parameters m and q in function of the parameters h (that simulates the persistent stimulus field),”alpha” (network’s storage capacity) and T (thermal noise), for the model on a fully connected network and symmetric connections. We analyze the recognition and storage capacity properties of the network, resulting on the phase diagram “alpha” x T for the model, showing the dependence of the recognition transition on the value of h. For the deterministic case (T = 0), we perform numerical simulations, where we develop and improve a computational algorithm using the multi-spin coding technique. Our simulation results show a good agreement with the analytical ones. All the results indicate na increase of the recognition capacity when increasing the h parameter, which controls the intensity of the stimulus field.