Comparison of growth models between artificial neural networks and nonlinear regression analysis in Cherry Valley ducks

The objective of this research was to compare growth models of Cherry Valley ducks. The multilayer perceptron (MLP) method and the radial basis function (RBF) method are both artificial neural networks that offer an alternative to nonlinear regression analyses (asymptotic exponential, logistic, cubic curvilinear, and Gompertz). To describe the growth curve, average BW was used. Training data consisted of alternate-day BW beginning with the first day, and validation data consisted of BW on other days. The R(2) and root mean square error of each model were determined for the training data. Mean absolute deviation and mean absolute percentage of error were used as the error measurements for the validation data. The RBF improved R(2) and root mean square error more than did the cubic curvilinear, Gompertz, logistic, asymptotic exponential, and MLP models. The error measurements of the Gompertz, RBF, and cubic curvilinear models were significantly lower than that of MLP. However, the mean absolute percentages of error of the asymptotic exponential, logistic, and MLP models were not significantly different from each other. It could be concluded that RBF produced more accurate predictions than MLP, but it did not produce more accurate predictions than the Gompertz and cubic curvilinear functions for estimating the BW of Cherry Valley ducks.