Mathematical description of growth curve in Kurdish sheep using artificial neural network and its comparison with non-linear models

The objective of this study was to compare artificial neural network (ANN) with non-linear models including Brody, Gompertz, Logistic and von Bertalanffy for predicting the growth curve of Kurdish sheep. The database comprised of 17659 body weights from birth to yearling of 5074 lambs belonging to 162 rams and 1968 ewes during 1996-2013. The ANN model was developed according to three-multilayer perceptron with five nodes in each layer, Sigmoid-Axon function and Levenberg-Marquat learning rule by Neuro Solution software. Non-linear models were analyzed by the NLIN procedure of SAS program. The goodness of fit of models and their comparisons were conducted by using the coefficient of determination (R2), residual mean square (MSE), root of the residual mean square (RMSE), mean absolute deviation (MAD), Akaike’s information criterion (AIC) and Bayesian information criterion information criterion (BIC). The influences of fixed effect on model parameters were analyzed on the optimum model. The results revealed that the ANN had the highest accuracy (r= 0.9735) and the lowest error (MSE= 0.9170, RMSE= 3.452, MAD= 2.424) and described the growth curve better than the other models. Among all non-linear models, the Brody model had the highest coefficient of determination (R2= 0.966) and the lowest AIC, BIC, MAD and RMSE values indicating the best fit for both sexes. Male lambs, single lambs and those gave birth in winter had the highest mature weight and growth rate. The evaluation criteria indicated that the ANN had a suitable potential to predict growth curve of Kurdish sheep, after that the Brody model fitted the data better than the other non-linear models.