2.8. Modelling using ANN

In this study, a commercial software NeuralPower version 2.5 (CPC-X Software) was used to predict the glycemic index and glycemic load of the formulated cookies. The ANN architecture included an input layer with two neurons (baking temperature and time), an output layer consisting of three neurons (RS, pGI and GL) as well as a hidden layer. To develop an optimal network topology for the model, the number of neurons, as well as the transfer function of hidden and output layers, were determined iteratively through the development of many networks. Multilayer full feedforward (MFFF), as well as multilayer normal feedforward (MNFF) neural networks, were used to predict the output variables, while the training of the data sets was done using different learning algorithms such as incremental backpropagation (IBP), quickprob (QP), genetic algorithm (GA), batch backpropagation (BBP), and Levenberg-Marquardt algorithm (LM). The experimental data obtained from the central composite design was split into two: training and testing data sets. The training dataset consists of 8 experimental runs while the remaining dataset was used for testing.