Training Networks

An important step while developing neural networks is to train them. Therefor training data is presented to the network and the error of the network is calculated. With this information the weights and thresholds in the network can be adjusted so that the net output matches the desired result.


The panel 'Training Networks' provides access to all functions to train networks of the JACVANN framework. Necessary parameters appear by selecting a training algorithm. Only Backpropagation is stable, Resilent Propagation and Quick Propagation are still experimental. If 'Init by Nguyen and Widrow' is deselected random initialisation at the initialisation range will be performed. By deselecting 'Init Net' initialisation could be disabled before training starts.

The 'Training Goals' section provides several abort criteria. 'Epochs' sets the maximum number of training epochs, 'Error' sets the error to reach by the test set. The error function could be selected at 'Error Type'. 'Bad Runs' sets the number of epochs in witch the test set results have to get better than the best reached result.

Before starting training a dataset has to be loaded. This is done by opening a .csd file containing the dataset of complex numbers to use. Therefor a promt is used:


Before starting training the amount of data from the data set should be used for testing. If 0.0 is entered, training goals were based on the training data error.


When training has reached a training goal the results were presented at a smal window. By deselecting 'Init Net' training can be continued at this state.


Next: Validate Network