Comparison of Multiple Linear Regression and Artificial Neural Network Models Goodness of Fit to Lactation Milk Yields

In this study, effects of lactation length, calving year and service period on lactation milk yield of Holstein Friesians were modeled with multiple regression and artificial neural networks (ANN) and compared goodness of fit of models. Analyses were carried on five lactations milk yields of 305 Holsteins calved at 2006, 2007 and 2008 years. After several experiments, hidden layer number was taken one and hidden nodes number were found three for the chosen architecture. Moreover, convergence criteria, maximum iteration number and epoch number were taken as 1.10(-6), 50 and 20, respectively. Adjusted coefficient of determination (R-2),root mean square error (RMSE), mean absolute deviation (MAD), mean absolute percentage error performance criteria (MAPE) were used for comparison of artificial neural network and multiple linear regression models goodness of fit. After analysis R-2 values were found among 0.62-0.85 for the five lactations with neural networks model. RMSE, MAD and MAPE criteria also were found among 480.9-1682.8, 325.2-1381.7 and 6.1-20.2, respectively. These criteria were found for R-2, RMSE, MAD and MAPE among 0.30-0.75, 1964.8-30008.7, 1576.6-2458.3 and 24.7-35.6, respectively for multiple linear regression. When the models were compared, artificial neural networks model gave better fit than multiple linear regression models. Consequently, artificial neural networks was determined an alternative method to multiple regression analysis.