Artificial neural network prediction of amino acid levels in feed ingredients

Abstract

Artificial Neural Networks (ANN), which are biologically inspired tools, serve as an alternative to regression analysis for complex data. Based on CP or proximate analysis (PA) of ingredients, two types of ANN and linear regression (LR) were evaluated for predicting amino acid levels in corn, wheat, soybean meal, meat and bone meal, and fish meal. The two ANN were a three layer Backpropagation network (BP3), and a General Regression Neural Network (GRNN). Methionine, TSAA, Lys, Thr, Tyr, Trp, and Arg were evaluated and R2 values calculated for each prediction method. Artificial neural network training was completed with NeuroShell 2 using Calibration to prevent overtraining. Ninety percent of the data were used as the input for the LR and the two ANN. The remaining 10% (randomly extracted data) were used to calibrate the performance of the ANN. As compared to LR, the R2 values were largest when PA input and GRNN were used. The BP3 did not consistently improve the R2 values for either CP or PA inputs as compared to LR. Each neural net can be incorporated into a computer or spreadsheet program.