Redes neurais artificiais e análise discriminante linear como alternativas para seleção entre famílias de cana-de-açúcar

• Autor(a) principal: Moreira, édimo Fernando Alves
• Data de Publicação: 2014
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: LOCUS Repositório Institucional da UFV
• Texto Completo: http://locus.ufv.br/handle/123456789/4075
• topic: Melhoramento de plantas; Saccharum spp.; Inteligência artificial; Plant breeding; Saccharum spp .; Artificial intelligence; CNPQ::CIENCIAS AGRARIAS

One of the challenges in breeding of cane sugar programs is the efficient selection of genotypes in the early stages. This challenge arises from the large number of genotypes evaluated and the operational difficulty of weighing all the main plots, required in the main selection methods. The objective of this study is to compare the modeling by artificial neural networks, Fisher s linear discriminant analysis and the selection of families above the overall average for the variable ton of cane per hectare estimated (TCHe) as alternatives for selecting promising families in cane sugar based on indirect character number of stalk (NC), stalk diameter (DC) and stalk height (AC). First was done the analysis for modelling for neural networks in 4 different scenarios (with simulation and standardization of variables; with simulation and without standardization of variables; without simulation and with standardization of variables; without simulation and without standardization of variables). The worst results occur in scenario 4, without standardization and without simulation and the best results occurred in scenario 1, where the variables were standardized and were simulated values of DC, NC, AC and TCHr to 1,000 families. Subsequently, the modeling was done through discriminant analysis in the best scenario, ie, that where there was simulation, and standardization of input variables. To evaluate the methods, neural networks, linear discriminant analysis and selection for TCHe, will be used the apparent error rate (TEA) and one error rate (TE1) obtained for matrix confusion. The simulation and standardization improve the performance of neural network models. The modeling via neural networks and discriminant analysis provide best results when compared to commonly used strategy, which is based on the estimation of the variable ton of cane per hectare. Comparing the models of neural networks with discriminant analysis, neural network gives better results.