Optimization methods for the definition of MLP neural network architectures and weights / Optimization methods for the definition of MLP neural network architectures and weights / MÃtodos de otimizaÃÃo para definiÃÃo de arquiteturas e pesos de redes neurais MLP / MÃtodos de otimizaÃÃo para definiÃÃo de arquiteturas e pesos de redes neurais MLP

AUTOR(ES)
DATA DE PUBLICAÇÃO

2005

RESUMO

This dissertation proposes modifications in the Yamazaki method for the simultaneous optimization of Multilayer Perceptron (MLP) network weights and architectures. The main objective is to propose a set of modifications with respective validations aimed at creating a more efficient optimization process. The optimization hybrid algorithm is based on the simulated annealing and tabu search algorithms as well as the Yamazaki method. The modifications are carried out in the implementation criteria, such as the neighbor generation mechanism, cooling schedule and cost function. One of the main points of this dissertation is the creation of a new neighbor generation mechanism aimed at increasing the search space. The cooling schedule is of great importance in the convergence of the algorithm. The cost of each solution is measured as the weighed average between the classification error for the training set and the percentage of connections used by the network. The databases used in the experiments are: odor classification proceeding from three vintages of wine and gases classification. The statistical substantiation for conclusions observed through results obtained by way of hypothesis tests. A study of the execution time was carried out separating the phases of global optimization from the phase of local refinement. It was concluded that with the new neighbor generation mechanism, the use of backpropagation was unnecessary, thereby obtaining a substantial gain in execution time. The optimization hybrid algorithm presented, to both the databases, the lowest value of the average classification error for the training set as well as the lowest amount of connections. Moreover, the execution time was reduced by an average of 46.72%

ASSUNTO(S)

yamazaki method reconhecimento de padrÃes otimizaÃÃo teste de hipÃteses ciencia da computacao arquiteturas de redes multilayer perceptron (mlp) redes neurais artificiais artificial noses optimization artificial neural networks narizes artificiais simulated annealing pattern recognition multilayer perceptron (mlp) network architectures metodologia yamazaki hypothesis tests simulated annealing tabu search tabu search

ACESSO AO ARTIGO

Documentos Relacionados