art_doc.txt

自适应共振神经网络可以实现自动学习与分类具有自适应的功能

		To get the input size, use GetNetInSize.
		To get the output size,use GetNetOutSize.

2.3.3  Training The Network

	The network is trained using the procedure, TrainSet.  TrainSet
	must be passed a pointer to an initialized network of type "netTYPE",
	a pointer to a set of training patterns of type "setTYPE", and a
	maximum number of epochs for training.

	Before calling TrainSet, network vigilance levels and recoding rates
	can be set using SetArtVigil, SetArtBeta and SetMapVigil.

	This procedure will continue to train the network until stable
	learning has been achieved or the maximum number of training epochs
	is reached.

	After each training epoch, TrainSet updates information about the
	number of resets and mismatches that occurred during that training
	epoch.  This information can be accessed using GetArtResets and
	GetNetMismatch.

	Additionally, TrainSet notes when the network has achieved stable
	learning.  This can be checked using GetNetDone, which returns TRUE
	when the network has achieved stable learning and FALSE otherwise.

2.3.4  Testing the Network

	The network's performance can be tested using TestNet which will
	return information about the number of ArtA and ArtB no-answers,
	the number of inputs correctly classified and the number incorrectly
	classified.  Note that for an ART network there is no such thing 
	as incorrectly classfying an input as the classification learning is 
	unsupervised.  The network can only make a mistake by not answering
	(ie, not meeting the vigilance level).

	Alternatively, the network can be tested by using ShowPat, which 
	will present a single pattern to the network from a pattern set.
	To access the resulting network activations, use the following:
	GetArtWinner returns the winning category for an Art component.
	GetMap returns the Mapfield activations.

2.3.5  Saving and Loading Networks

	All networks are stored in files with the extention ".net".  This
	extension is automatically added to the filename, when files are
	saved and loaded.

	To load or save a network use LoadNet and SaveNet without the ".net"
	extension.

2.3.6  Freeing Network Space

	When a network is no longer going to be used it is a good idea to
	clear the memory being used by the network before loading or
	initializing a new network.  This can be done using FreeNet.

	Failure to do so may result in problems with running out of memory.

2.4   Pattern Sets

	Patterns sets consits of a list of input (and in some cases output)
	patterns to be presented to the network, as well as specifics about
	the type of data contained in the pattern set.  They are stored in a
	data structure of type "setTYPE", and can be pointed to by a pointer
	of type "setPTR" both of which are defined in Art_Lib.h.

2.4.1  Creating a Pattern Set

	Pattern sets are created using MakeSet.  MakeSet creates a new pattern
	set given the desired number of input and output patterns for the
	pattern set as well as the type of patterns and two files containing
	the input and output data.

	Input and output patterns can be of type ANALOG or BINARY.
	Output patterns can also be of type NONE if the set is to contain
	only input patterns.

	The two data files containing input and output data should consist of
	one number per line, with no blank lines between patterns.  Two
	sample files "analog.dat" and "binary.dat" have been provided to
	demonstrate how this might be done.

	Patterns can be added to the pattern set using AddPattern.

2.4.2  Data "Type"

	Input and output data come in three types, NONE, BINARY and ANALOG.

	BINARY patterns must be either 0 or 1, and can be used with either
	ART1 or FUZZYART components.

	ANALOG input patterns must be between 0 and 1, and can only be used by
	FUZZYART components.

	An ART-type network doesn't use output patterns, so outputs can be
	of any type, including NONE.

2.4.3  Examining Pattern Set Parameters and Data

	Data in the pattern set can (and should) be accessed without reference
	to parts of the data structure itself.  A number of functions have
	been provided for easy access to pattern set information.

	The size and type of the input patterns for a pattern set can be
	checked using GetSetInSize, GetSetOutSize, GetSetInType, and
	GetSetOutType.

	Particular input and output pattern data can be retrieved using
	GetSetInput and GetSetOutput.

	To find out the number of patterns in the pattern set use
	GetNumPatterns.

2.4.4  Saving and Loading Pattern Sets

	All pattern sets are stored in files with the extention ".set".
	This extension is automatically added to the filename, when files are
	saved and loaded.

	Pattern sets can be loaded and saved using SaveSet and LoadSet
	without the ".set" extension.

2.4.5  Merging Pattern Sets

	Two pattern sets can be merged using MargeSets.

2.4.6  Freeing Pattern Set Space

	When a pattern set is no longer going to be used it is a good idea
	to clear the memory being used by the pattern set before loading or
	initializing a new pattern set.  This can be done using FreeSet.

	Failure to do so may result in problems with running out of memory.

2.5  Note on Error Checking and Variable Initialization

	The procedures in the Art Gallery attempt to do extensive error
	checking to ensure that all variables are properly initialized before
	they are used.

	GetNetInit and GetSetInit can also be used to check the initialization
	status of a Pattern Set or Network.

	Note, however, that there is always a small chance the the the area of
	memory choosen for the initialization marker will coincidentally already
	have the value of TRUE.  This is especially true on the PC platform.
	In this case initialization checks will fail GetNetInit and GetSetInit
	will return incorrect values.
	The only way to make 100% sure that procedures will return their
	expected values is to only call them with initialized pattern sets and
	networks.

2.6  Compliment Coding and Network Stability

	It is generally a good idea to use compliment coding when creating
	a new network to ensure that stable learning will occur.  Without
	compliment coding established categories can erode.  This is
	especially important when type NONE is selected for the ARTB network.
	When compliment coding is used the size of network inputs must be
	twice the size inputs in the pattern set.  The compliment coded
	values are automatically calculated by the program.

	When a network uses compliment coding, a 'negative' copy of the
	inputs is also presented to the network.  So, for example, if an
	input pattern consists of 5 values and compliment coding is used,
	the network must have an input size of 10:

	Pattern Set Input in pattern file:
		0.2 0.3 0.5 0.7 0.1

	Input presented to network when COMPLIMENT style is chose:
		0.2 0.3 0.5 0.7 0.1 0.8 0.7 0.5 0.3 0.9

___________________________________________________________________________
3.0     COMPILING THE ART GALLERY AS A DYNAMIC LINKED LIBRARY (DLL)
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
	The ART Gallery can be compiled as a dynamic linked using
	Borland C++.  This allows the ART Gallery functions to be called
	from other Window's applications.

	The DLL can then be by typing at the PC prompt:
		'make -f art_gal'

	The resulting DLL file will be callled:
		'Art_Gal.dll'


3.1  Using the ART Gallery DLL's in Visual Basic

	Once The ART Gallery has been compiled as a DLL, and declared in
	Visual Basic, the ART Gallery functions can be called from Visual
	Basic.

	The file 'Default.bas' contains a Visual Basic module demonstrating
	how the ART Gallery DLL's should be declared in Visual Basic.

	See section 4.2 for an example of how The ART Gallery can be used
	in Visual Basic.

___________________________________________________________________________
4.0     SAMPLE SIMULATORS
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
4.1   Unix and Dos Sample Text Simulator

	The file "Art_Sim.c" should have been included with the Art Gallery.
	It contains a sample of how the Art Gallery might be used to create
	a simple text driven neural network simulator.

	A sample training and testing set have also been provided. (See
	section 5.0).

	On Unix:
		Art_Sim.c can be compiled with the included Makefile by
		typing 'make' at the Unix prompt.

	On Dos: (Using Borland C++)
		Art_Sim.c can then be compiled by typing at the PC prompt:
		'make -f art_sim'

4.2   Sample Windows ART Gallery Simulator

	The file "win_gal.zip" contains the install program for an graphical
	windows version of The Art Gallery, 'Art_Gal.exe' created using The 
	Art Gallery functions complied as a Dynamic Linked Library.

___________________________________________________________________________
5.0     SAMPLE TRAINING AND TESTING FILES
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
	A sample training and testing pattern set file (train.pat and
	test.pat) should have been included with the Art Library.

	Each consists of 11 ANALOG input vectors and 6 BINARY output
	vectors.

5.1   Sample Networks for training set

	This data set can be used with many different types of networks.
	A couple examples are shown below:

	1) Unsupervised Learning (output vectors unused)
	     Network Type: ART
	     Art Type:     FuzzyART   *
	     Input Size:   11
	     Input Style:  NONE

	2) Supervised Learning with ArtB Module and Compliment Coding
	     Network Type: ARTMAP
	     ArtA Type:    FuzzyART   *
	     Input Size:   22         **
	     Input Style:  COMPLIMENT
	     ArtB Type:    ART1       ***
	     Input Size:   6
	     Input Style:  NONE

	3) Supervised Learning without training ArtB Module
	     Network Type: ARTMAP
	     ArtA Type:    FuzzyART   *
	     Input Size:   11
	     Input Style:  NONE
	     ArtB Type:    NONE
	     Input Size:   6          ****

	   * - Input must be of type FuzzyART since inputs are ANALOG
	  ** - When compliment coding input size is double the size
		 of the input vector in the pattern set
	 *** - Input type for the ARTB network can be either FuzzyART
		 of ART1, since both can take BINARY pattenrs
	**** - ARTB input size is the size of the input vector (not
		 zero) even when no ARTB training is used.

5.2   Sample Data for Making a Network