Sigmoid Activation Function in Selecting the Best Model of Artificial Neural Networks

Abstract

Abstract The aim of the research is to make predictions from the best architectural model of backpropagation neural networks. In determining the outcome in the form of a prediction model, the activation function in the artificial neural network is useful to transform an input into a certain output. In this study the activation function used is sigmoid. The study uses the case of population density in Indonesia considering that for developing countries population growth has many negative impacts such as increased poverty, unemployment and crime rates. This study uses data from Badan Pusat Statistic Indonesia for population categories in 2003-2015. The process of determining the architectural model is carried out 2 stages including: the training stage and the testing stage. The two architectural models used (5-2-1 and 5-10-1), the selection of architectural models is done by looking at several assessment parameters such as epoch, accuracy, MSE training and MSE testing. The process of training and testing of data is done by using the help of the Matlab application. The results of the study obtained that the architectural model 5-2-1 is the best model for predicting with an accuracy of 65%, MSE Training 0,0009997254, MSE Testing 0.0014897214 and Epoch 28548.