Redes neurais artificiais na discriminação de populações de retrocruzamento com diferentes graus de similaridade

• Ano de defesa: 2014
• Autor(a) principal: Sant'anna, Isabela de Castro
• Orientador(a): Cruz, Cosme Damião
• Banca de defesa: Nascimento, Moysés
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Viçosa
• Programa de Pós-Graduação: Mestrado em Genética e Melhoramento
• Departamento: Genética animal; Genética molecular e de microrganismos; Genética quantitativa; Genética vegetal; Me
• País: BR
• Palavras-chave em Português: Melhoramento genético; Análise discriminante; Inteligência artificial; Redes neurais
• Palavras-chave em Inglês: Breeding; Discriminant analysis; Artificial intelligence; Neural networks
• Área do conhecimento CNPq: CNPQ::CIENCIAS BIOLOGICAS::GENETICA::GENETICA QUANTITATIVA
• Link de acesso: http://locus.ufv.br/handle/123456789/4801

The correct classification of individuals has a top importance for the genetic variability preservation as well as to maximize gains. The multivariate statistical techniques commonly used in these situations are the Fisher and Anderson discriminant functions, allowing to allocate an initially unknown individual in a probably g population or predefined groups. However, for higher levels of similarity such as backcross populations these methods has proved to be inefficient. Currently, much has been Said about a new paradigm of computing, artificial neural networks, which can be used to solve many statistical problems as similar subjects grouping, time-series forecasting and in particular, the classification problems. The aim of this study was to conduct a comparative study between the Fisher and Anderson discriminant functions and artificial neural networks through the number of incorrect classifications of individuals known to belong to different simulated backcross with increasing levels of populations similarity. The dissimilarity, measured by Mahalanobis distance, was a concept of fundamental importance in the use of discrimination techniques, due the quantification of how much populations were divergent. Data collection was done through simulation using the software Genes. Each population generated was characterized by a set of elements measured by characteristics of a continuous distribution. The relations of relatives and hierarchical structuring were established considering genetically divergent populations, F1 hybrid and five generations of backcrossing in relation to each of the relatives, establishing measures of effectiveness of the tested methodologies. The phenotypic data of populations were used to establish the Fisher and Anderson discriminant function and the calculation of the apparent error rate (AER), which measures the number of incorrect classifications. The ERA Estimations were compared with those obtained by means of neural networks. The artificial neural network is shown as a promising technique to solve classification problems, once it had a number of incorrect individuals classifications smaller than the data obtained by the discriminant functions.