Application of Artificial Neural Network and Support Vector Machines in Predicting Metabolizable Energy in Compound Feeds for Pigs

Hamed Ahmadi, Markus Rodehutscord

In the nutrition literature, there are several reports on the use of artificial neural network (ANN) and multiple linear regression (MLR) approaches for predicting feed composition and nutritive value, while the use of support vector machines (SVM) method as a new alternative approach to MLR and ANN models is still not fully investigated. The MLR, ANN, and SVM models were developed to predict metabolizable energy (ME) content of compound feeds for pigs based on the German energy evaluation system from analyzed contents of crude protein (CP), ether extract (EE), crude fiber (CF), and starch. A total of 290 datasets from standardized digestibility studies with compound feeds was provided from several institutions and published papers, and ME was calculated thereon. Accuracy and precision of developed models were evaluated, given their produced prediction values. The results revealed that the developed ANN [R(2) = 0.95; root mean square error (RMSE) = 0.19 MJ/kg of dry matter] and SVM (R(2) = 0.95; RMSE = 0.21 MJ/kg of dry matter) models produced better prediction values in estimating ME in compound feed than those produced by conventional MLR (R(2) = 0.89; RMSE = 0.27 MJ/kg of dry matter). The developed ANN and SVM models produced better prediction values in estimating ME in compound feed than those produced by conventional MLR; however, there were not obvious differences between performance of ANN and SVM models. Thus, SVM model may also be considered as a promising tool for modeling the relationship between chemical composition and ME of compound feeds for pigs. To provide the readers and nutritionist with the easy and rapid tool, an Excel(®) calculator, namely, SVM_ME_pig, was created to predict the metabolizable energy values in compound feeds for pigs using developed support vector machine model.