Improving neural network prediction of amino acid levels in feed ingredients.

摘要: Abstract Artificial neural networks (ANN) were trained to predict the amino acid (AA) profile of feed ingredients. The ANN more effectively identified complex relationship between nutrients and ingredients than linear regression (LR). Three types (NeuroShell 2): three-layer backpropagation (BP3), Ward Backpropagation (WBP), a general network (GRNN); LR (SAS Proc GLM) used AA level in corn, soybean meal, meat bone fish wheat based on proximate analysis. In contrast past study, variety alternative training parameters examined improve performance. Predictive performance was judged basis maximum R2 value resulting from all defaults tested. Advanced selection led further improvement performance, especially within GRNN architecture. 34 35 developed, for each individual ingredient higher LR, BP3, or WBP prediction methods. For example, highest Met corn 0.32 0.40 3LBP, 0.51 WBP, 0.95 also improved overall as compared results previous study. values (GRNN) Met, TSAA, Trp were: 0.78, 0.81 0.44, previously, 0.95, 0.96 0.88, current Current meal R 0.92; 0.94; Lys, 0.90. mean 0.97; 0.97. computation is successful statistical analysis predicting levels