Comparison of different activation functions in artificial neural networks for prediction of lactation milk yield

This study aims to estimate lactation efficiency in dairy cattle using an artificial neural network. The dataset used in the study was obtained from a private farm operating in the Thrace region. In the modeling process, both Min-Max and Z-Score normalization techniques were applied due to the different scales of the data; thus, the variables were made comparable. In order to evaluate the model performance reliably, a 10-fold cross-validation method was used. Within the scope of the study, various activation functions such as Sigmoid, Tanh, ReLU, ELU, Leaky ReLU, Softplus and Swish were tried; single hidden layer ANN models containing increasing numbers of neurons from 2 to 100 were created for each activation-normalization combination. Model performances were evaluated with criteria such as R², MAE, RMSE, MSE and MAPE. According to the results obtained, the model that reached the highest accuracy value in the test data was the 55-neuron structure created using Min-Max normalization and ReLU activation function. The performances of this model on the test data were calculated as R² = 0.9540, MAE = 0.0265, RMSE = 0.0376 and MAPE = 6.88. In the study, estimation was also made with the MLR model. The best test performance of the MLR model was obtained with the data set where Min-Max normalization was applied; R² = 0.9162, MAE = 0.0359, RMSE = 0.0376 and MAPE = 6.883 were determined. These findings show that the ANN model provides lower error rates and higher estimation accuracy compared to the MLR model, especially when optimized with the appropriate activation function and neuron number selection.