New Features of the NeuroCoM Release 1.2
The following features have been added to the NeuroCom 1.2:
- 3D View
The transfer function of the trained network is displayed by a
3D diagram.
- Toolbar
Some additional menu commands are available by the toolbar.
- File formats
The file format settings dialog is accessable from every dialog box where
a learning file or another data input or output file has to be specified.
The output file format may be switched from the "Data" mode
(= learning file format) to the "Comparison" mode (comparison of desired and
calculated values).
- Statistics
A new "Statistics" function is available to perform a quick analysis
of the average and the maximum error which occurs when a test data file is
evaluated by the trained network.
- Training methods
Some additional training methods and options are implemented now:
- Backprop: The "offline mode" which was already used in previous
versions of the NeuroCoM. The weights are modified always after processing one
training pattern.
- Adaptive Backprop: Using this "batch mode", the weights will be
modified only after processing the entire training data file. An individual
learning rate for each weight will be adapted according to the weight changes
of successive training steps (continuous or alternating sign of the changes).
- Quickprop: This algorithm tries to calculate the minimum of
the local error curve and to modify the weights accordingly.
- Flat Spot Elimination: Using this option, changes of the input
weights become more effective if the current operating point of the neuron is
located within the saturation area of the sigmoid function.
- Output Error Distortion: This performs a non-linear distortion
of the calculated error of an output neuron. In this way, the influence of
larger errors will be increased.
- Online documentation
The greatest part of the manual is available by the online help now.
Softcomputing ·
Contents