Submitted by Lara Oliveira (lara@ufersa.edu.br) on 2017-09-08T22:30:39Z No. of bitstreams: 1 AldemárioAS_DISSERT.pdf: 18856416 bytes, checksum: dcd37bbe9d111ef051c4d27c3481a41f (MD5) / Approved for entry into archive by Vanessa Christiane (referencia@ufersa.edu.br) on 2017-09-11T16:27:51Z (GMT) No. of bitstreams: 1 AldemárioAS_DISSERT.pdf: 18856416 bytes, checksum: dcd37bbe9d111ef051c4d27c3481a41f (MD5) / Approved for entry into archive by Vanessa Christiane (referencia@ufersa.edu.br) on 2017-09-11T16:28:25Z (GMT) No. of bitstreams: 1 AldemárioAS_DISSERT.pdf: 18856416 bytes, checksum: dcd37bbe9d111ef051c4d27c3481a41f (MD5) / Made available in DSpace on 2017-09-11T16:29:16Z (GMT). No. of bitstreams: 1 AldemárioAS_DISSERT.pdf: 18856416 bytes, checksum: dcd37bbe9d111ef051c4d27c3481a41f (MD5) Previous issue date: 2017-08-18 / The training of Multilayer Perceptron Neural Network (MLPNN) done by exact algorithm to find the maximum accuracy is NP-hard. Thus, we use the algorithm Back-Propagation who needs a starting point (weights and bias initials) to compute the training of the MLPNN. This research has developed and implemented a heuristic algorithm HeCI - Heuristic to Calculate Weights and Bias Initials - to compute the data to train the MLPNN and return the starting point for the Back-Propagation. HeCI uses Principal Component Analysis, Least Square Method, Probability Density Function of the Normal Gaussian Distribution, two strategic configurations, and partially controls the number of MLPNN training epochs. Experimentally, HeCI was used with Back-Propagation in MLPNN training to recognize patterns and solve data classification problems. Six case studies with datasets between Health, Business and Botany were used in the experiments. The methodology of this research uses Deductive analysis by the Experimental method with Quantitative approach and hypothesis tests: Test of Fridman with post Teste of Tukey HSD Post-hoc and Wilcoxon Test-M W. The results of accuracy have increased significantly improving attested by evaluation of tests of hypotheses, inferringstatistical robustness of the result motivated by HeCI / O treinamento de Rede Neural Perceptron Multicamadas (RNPM) feito por algoritmo exato para encontrar a máxima acurácia é NP-Difícil. Sendo assim, usa-se o algoritmo "Back-Propagation" que necessita de um ponto de partida (pesos e bias iniciais) para computar o treinamento da RNPM. Esta pesquisa desenvolveu e aplicou um algoritmo heurístico HeCI - Heurística para Calcular Pesos e Bias Iniciais - para computar os dados de treinamento da RNPM e retornar o ponto de partida para o "Back-Propagation". A HeCI usa Análise de Componentes Principais, Método dos Mínimos Quadrados, Função de Densidade de Probabilidade da Normal Distribuição Gaussiana, duas configurações estratégicas e controla parcialmente o número de épocas de treinamento da RNPM. Experimentalmente, a RNPM foi treinada usando "Back-Propagation" com HeCI, para reconhecer padrões e resolver problemas de classificação de dados. Seis estudos de caso com "datasets" entre as áreas de Saúde, Negócio e Botânica foram usados nos experimentos. A metodologia desta pesquisa usa análise Dedutiva pelo método Experimental com abordagem Quantitativa e testes de hipóteses: Teste de Fridman com Pós Teste de Tukey HSD Post-hoc e Teste de Wilcoxon-M-W. Os resultados de acurácia incrementaram melhoria significativa atestada pela avaliação dos testes de hipóteses, inferindo estatisticamente robustez de resultado motivado pela HeCI / 2017-09-08