Comparison of artificial neural network performance with some mathematical functions in predicting of lactation curve of Iranian dairy cows

In order to compare the performance of neural network with some mathematical functions for predicting of lactation curve of Iranian Holstein dairy cows, a total of 1,085,525 milk test day records from first-parity dairy cows calved during 1983-2012 were used. Fitting the lactation curve was performed by brnn package (for neural network) and also by some mathematical functions (including Wood, Wilmink, Ali-Schaeffer and Pollott-Gootwine) using R software based upon average milk yield, fat and protein percentage test day records. The criteria of AIC, BIC, RMSE and adjusted R2 were utilized to evaluate goodness of fit. The results showed that the Bayesian neural network (brnn) had a better fit than mathematical functions in describing the standard curve shape of Iranian Holstein dairy cows. Among the mathematical functions used for milk yield, Wilmink model had a better fit while for milk fat percentage and milk protein percentage, Ali-Schaeffer model showed a better fit performance. Therefore, it could be suggested that brnn is an appropriate option to be applied to fit the lactation curve of Iranian Holstein dairy cows.