Prediction of amino acid profiles in feed ingredients:: Genetic algorithm calibration of artificial neural networks

Linear regression (LR) has been used to predict the amino acid (AA) profiles of feed ingredients, given proximate analysis (PA) input. Artificial neural networks (ANN) have also been trained to predict AA levels, generally with better results. Past projects have indicated that ANN more effectively identified the complex relationship between nutrients and feed ingredients than did LR. It was shown that the maximum R2 value, a measurement of the amount of variability explained by the model, was highest when a general regression neural network (GRNN) with iterative calibration (GRNNIT) was used to train the ANN. This was in comparison to LR, Ward backpropagation (WBP) or 3-layer backpropagation (3BP) architectures. The current study investigated the potential of a new, advanced method of calibration using the genetic algorithm (GA) to optimize GRNN smoothing values. Calibration of an ANN allows the neural network to generalize well and therefore provide good results on new data. A GRNN architecture (NeuroShell 2® Software) with GA calibration (GRNNGA) was used to train an ANN to predict AA levels in maize, soya bean meal (SBM), meat and bone meal, fish meal and wheat, based on proximate analysis input. Within the GRNNGA architecture, ANN were trained with either an Euclidean or City Block distance metric and a (0,1), (−1,1), (logistic) or (tanh) input scale. Predictive performance was judged on the basis of the maximum R2 value. In general, maximum R2 values were higher when the GA calibration was used in comparison to LR. For example, the highest methionine (MET) R2 value for SBM was 0.54 (LR), 0.81 (3BP), 0.87 (WBP), 0.92 (GRNNIT) and 0.98 (GRNNGA). Genetic algorithm calibration of GRNN architecture led to further improvements in ANN performance for AA level predictions in most of the cases studied. Exceptions were the TSAA level in SBM (0.94 with GRNNIT vs. 0.90 with GRNNGA) and the TRY level in maize (0.88 with GRNNIT vs. 0.61 with GRNNGA).