Evaluation of Normalization Techniques on Neural Networks for the Prediction of 305-Day Milk Yield

Abstract

In this study, the impact of data preprocessing on the prediction of 305-day milk yield using neural networks were investigated with regard to the effect of different normalization techniques. Eight normalization techniques “Z-Score, Min-Max, D-Min-Max, Median, Sigmoid, Decimal Scaling, Median and MAD, TanhEstimators" and five different back propagation algorithms “Levenberg-Marquardt (LM), Bayesian Regularization (BR), Scaled Conjugate Gradient (SCG), Conjugate Gradient Back propagation with Powell-Beale Restarts (CGB) and Brayde Fletcher Gold Farlo Shanno Quasi Newton Back propagation (BFG)” were examined and tested comparatively for the analysis. Neural network architecture was optimized and tested with several experiments. Results of the analysis show that applying different normalization techniques affect the performance and the distribution of outputs influences the learning process of the neural network. The magnitude of the effects varied with the type of back propagation algorithms, activation functions, and network's architectural structure. According to the results of the analysis, the most successful performance value in the 305-day milk yield estimation was obtained by using the neural network structured by using the Decimal Scaling normalization technique with the Bayesian Regulation algorithm (R2Adj = 0.8181, RMSE= 0.0068, MAPE= 160.42 for test set; R2Adj =0.8141, RMSE= 0.0067, MAPE= 114.12 for validation set).