------------------------------------------------------
QwikNet revision history
Copyright (C)1996-1998 Craig Jensen
------------------------------------------------------

Changes for version 2.20
------------------------
Added the ability to output network code in the Delphi Pascal programming language. Special thanks to Mr. Magnus Jansson for his help on this feature.

Added the ability to save networks as Matlab M-files.

Added several additional stopping criteria and the ability to monitor the testing data file for stopping.

Corrected several internal resource leaks.

Implemented a new automatic scaling feature that eliminates the need for users to manually scale their training and testing data files.

Changes for version 2.15
------------------------
Optimized the startup code. QwikNet now loads considerably faster.

Fixed the problem of failing to find the correct help file from the main window that sometimes occurred.

Fixed the kernel fault that sometimes occurs when attempting to load large files with the shareware version.

Added several new example files.

Changes for version 2.14
------------------------
Several minor bug fixes.

Changes for version 2.13
------------------------
Added ability to display error statistics for the testing data during training.

Changed the training error plot so that it displays the RMS training, testing and cross-validation error.

Changes for version 2.12
------------------------
Converted all routines from single precision floating point to double precision. This will provide more accurate and faster training.

Fixed bug in the Network analysis plot that caused the plot to continuously redraw itself.

Added max absolute error to the training info.

Changes for version 2.1
-----------------------
Added the ability to choose between sigmoid, tanh, Gaussian, and linear activation functions for each layer.

Added weight decay and input noise to learning algorithms. This can help produce networks with better generalization capabilities.

Added axis control to the 3D and error analysis plots.

Improved redraw speed of 3D plot.

Fixed tabbing sequence for all windows.

Added auto axis control to error analysis plots.

Added ability to analyze training data, testing data, or both for all error analysis plots.

Several improvements to main interface.

Weights are now randomized in the range (-2.4/Fi,+2.4/Fi), where Fi is the Fan In or number of weights into that particular neuron. This is a popular heuristic and decreases the chances of the network becoming prematurely saturated.

Added ability to write network as C code. This makes it easier to make use of networks trained by QwikNet in custom applications.

Training and testing data files may now begin with any number of comment lines. This makes it easier to keep track of your data files.