Máquina de aprendizado extremo para predição de ganho médio diário à desmama em bovinos

Abstract

Extreme Learning Machines, Extreme Machine Learning (EML/IGASE), unlike other Artificial Neural Network (ANN) training algorithms that adjust network parameters through the iterative presentation of training patterns, include intermediate neuron numbers and perform projection random in the hidden layer, in general of high dimensionality in the complex connections with the other variables of direct inputs, the weights are selected in a random way injected and bijected in the complex composite functions, without need of training. The over- sizing of the EML/IGASE becomes necessary for the smoothing of the response, guarantees the capacity of generalization and transfers of complex functions in the temporal series of phenotypic expressions of the animals. The objective of this work is to predict the average weight gain at weaning (GMPD) phenotype as a function of direct variables measured in animals within and between seasons (SF) and farms (FAZ), using EML/IGASE. 8,812 progeny records were used for 11 harvests, descendants of 272 bulls, from 09 farms located in different states of the federation, Mato Grosso (MT), Mato Grosso do Sul (MS), Paraná (PR) and São Paulo (SP). ). The phenotypes studied were weight at weaning (PD), average weight gain at weaning (GMPD), conformation at weaning (CPD), musc ulature at weaning (MPD) and early weaning (PPD). Extreme Machine Learning (EML/IGASE) was able to predict the mean weight gain at weaning (GMPD) phenotype, based on the other variables measured in the animals within and between seasons and farms studied, with low Mean Squared Error (MSE) in the robust interactions and transfers of complex functions from multi-input neural architectures and different levels of environmental attributables to the expressed phenotypes. In the univariate analyzes with the input variable weights at weaning (PD) transferring to the average weight gain at weaning (GMPD) the EML/IGASE performed complex numeric al predictions in 08 productive harvests of the 09 farms, where the Mean Squared Errors (EQM) between predicted and actual variables of GMPD ranged from 0.09 to 13.96%. For the multivariate analyzes with the four input variables weight at weaning (PD), conformation at weaning (CPD), musc ulature at weaning (MPD) and precocity at weaning (PPD) in 09 harvests of the 09 productive farms, the Mean Squared Errors (NDE) between predicted and actual GMPD variables ranged from 0.08 to 26.30%. The complex transfer functions were robust to noise tests up to the second decimal place in the numeric al predictions considered in the study.