Welcome to the new version of NEFCLASS-PC. Starting with version 2.0
we will supply this whats.new file describing the new features of
the program.

============================================================================
Version 2.04, Release Date: February, 1, 1996

This is the fourth bug fix to version 2.0

* When a network was saved to a file before it was trained, it could
  not be read again to use it for training. Now a network that is
  read from file can also be correctly trained, if it is uninitialized, 
  and if it has no rules.

* When a network was saved to a file, then the fuzzy set parameters
  were formatted incorrectly, when the domain boundaries of a variable 
  were greater then 100. Such a network file could not be read again.
  The parameters are now formatted correctly.

* After the INRANGES statement in the pattern file only integer values
  could be specified. Now real values are accepted.

* If a network without rules is read from a file, then rule learning
  will be invoked when the training process is started.

* NEFCLASS is now compiled with the compiler of Borland Pascal 7.01
  (the release before used the 7.0 version). There are new versions 
  of Borland's redistributables rtm.exe and dpmi16bi.ovl included in
  this distribution.

============================================================================
Version 2.03, Release Date: January, 15, 1996

This is the third bug fix to version 2.0
This version was not officially released

* If a network was saved directly after it was created, such that
  it had no rules at this time, and the domains of the input variables
  were not set by a pattern file during training, then it could not be 
  used for training after loading it again.  NEFCLASS did not recognize
  the incorrect domains, and did not use the ranges from the pattern file
  to initialize the re-loaded network properly. This was corrected.

* NEFCLASS did not keep any rules with "negative" performance, when it
  evaluated a rule base for "best" or "best per class" learning. A rule
  has a negative performance, if its activation accumulated over all 
  patterns for its own class (determined by the conclusion) is less then
  then the activation accumulated over all patterns and all other classes.
  NEFCLASS now also accepts rules with negative performance, because this
  seems to be necessary for hard problems.
  If you do not want these kind of rules, you must delete them manually
  from the rule base, after you have identified them by looking into the
  NEFCLASS.PRO protocol file.
  A later version of NEFCLASS will have a switch for this.

============================================================================
Version 2.02, Release Date: November, 12, 1995

This is the second bug fix to version 2.0

* When a saved network has rules in which some variables are not used, the
  program did not load it correctly. This was fixed.
* The Change Directory Dialog in the File Menu does now work.

============================================================================
Version 2.01, Release Date: November, 2, 1995

This is a bug fix to version 2.0

* The scroll bar of the antecedent list in the rule editor was fixed
* The program crashed, when patterns without class information were used
  to test a network. This was fixed.

============================================================================
Version 2.0, Release Date: October, 8, 1995

* NEFCLASS-PC does now run in protected mode, and therefore needs an 80286 
  processor. It can make use of extended memory.
* A network can now have 50 input/output units, and up to 500 rule units.
* You can specify variable and class names when you create a new network.
* Each input variable can have an individual number of fuzzy sets.
* You can enter a description for a network.
* You can manually enter, delete and modifiy rules represented within a
  network.
* A network can be changed after it was created or trained.
* If a new network is specified, an eventually loaded pattern set will
  be deleted from memory.
* The number of patterns that can be loaded does only depend on the
  available memory.
* A pattern set can have a description (name).
* The learning parameters have been split in learning control parameters
  and learning parameters.
* There are two new constraints for learning membership functions
  (no passing, asymmetrical learning).
* All 3 rule learning procedures do now work.
* Rule learning is now independent from the sequence of patterns.
* The rule base can be relearned.
* A new dialog computes statistics of the pattern set, and can graphically 
  display bar charts of a feature, and 2-dimensional projections of the 
  patterns. These graphics can be plotted to an HPGL file.
* An entry was added to the Help menu that informs about the memory
  situation.
* There is now a warning before leaving the program, if a network
  was specified.
* NEFCLASS-PC looks for the environment variable BGIPATH to find the
  graphic drivers (.bgi and .chr files). If this variable is not set, it
  expects these files to be in the current directory.
* The online help was extended.

Known Bugs in Version 2.0:

- though memory management was improved a little bit (the program
  will not crash, when you try to load a pattern set that does not
  fit into memory), it will still crash, if memory is low, and certain
  dialogs are invoked.

============================================================================
Version 1.5, Release Date: February 22, 1995

This was the first released version. Prior versions were used for
internal validation only.

============================================================================