Redes neurais na predição de valores energéticos de alimentos para aves

Abstract

The precise knowledge of chemical composition and food energy values is necessary to formulate nutritionally balanced rations.The energetic content may be determined using bioassays, nutritional composition tables and prediction equations based on food chemical composition. Recently, artificial neural network (ANN) models have been implemented to predict metabolizable energy of poultry feddstuffs. However, neither a committee of networks nor the uncertainties related to the predictions of the metabolizable energy were considered. In this work, an ANN model was implemented to predict the nitrogen-corrected apparent metabolizable energy values (AMEn) of feedstuffs for broilers. One Excel macro was created based on the results of the best ANN model, which allows an efficient way to predict the AMEn values. Another objective of this work was to quantify the uncertainty related to the predicted values of AMEn, considering Highest Probability Density (HPD) credibility intervals obtained by using a committee of neural networks. A test dataset, which was obtained from bioassays, was used to evaluate the percentage of correct predictions (PCP) of different feedstuffs. The PCP was used as quantitative criterion to compare the predictions obtained using the network committee and the multiple linear regression models. Furthermore, the accuracy of the models was evaluated based on their values of error measures between the observed and predicted values. The results of this study demonstrated that, the accuracy of the predicted values of AMEn by considering the mode of an empirical distribution, obtained from 1000 networks of a committee, was higher than that obtained by using the regression model. Thus, the network committee is recommended as the most accurate model to predict the AMEn values of the feedstuffs energetic concentrates used by the poultry feed industry.