brain12.doc

本例是一个神经网络仿真应用程序

	      The Brain
	      Version 1.2s

	      User's Manual
	      _____________________________










	      (c) Copyright, All Rights Reserved, 1994, DP Computing

	      DP Computing
	      PO Box 712
	      Noarlunga Center SA 5168
	      Australia

	      Internet:
		[e-mail]
		[e-mail]


      The Brain v1.2 - User's Manual                                Page 2
      ____________________________________________________________________



			       Table Of Contents
			       -----------------


		Introduction  .............................  3

		Program Files  ............................  4

		Introduction To Neural Networks  ..........  5

		   What Are Neural Networks?  .............  5

		   How Do Neural Networks Learn?  .........  5

		   Uses of Neural Networks  ...............  7

		Training And Testing The Brain  ...........  8

		   Training The Network  ..................  8

		   Hints On Training A Network  ...........  9

		   Testing The Network  ...................  9

		   Input File Layout (with examples)  ..... 10

		Tutorial  ................................. 14

		General Texts On Neural Networks  ......... 17

		License Agreement  ........................ 18

		Support Policy  ........................... 19

		Distribution Policy  ...................... 19

		About This Manual  ........................ 20

      The Brain v1.2 - User's Manual                                Page 3
      ____________________________________________________________________

   
				  Introduction
				  ------------

      Get ready to explore the exciting world of artificial intelligence.
      The Brain is an advanced neural network simulator that is simple
      enough to be used by non-technical people, yet sophisticated enough
      for serious research work.  Based upon the backpropagation learning
      algorithm, The Brain allows you to train the computer to learn what
      you want it to learn.  The Brain gives you a glimpse into the future
      of computing.

      With The Brain, you can create, train, and test your own neural
      networks.  Three sample networks have been included with this
      distribution package:

	 - a network to recognise the numerals 1, 2, and 3.

	 - a network to process the logical AND function.

	 - a network to process the logical XOR function.

      This manual will outline the features and capabilities of The Brain
      as well as providing you with a brief overview of the concepts and
      applications of neural networks.  Several excellent books are listed
      at the end of this manual for those interested in a more thorough
      introduction to neural networks.

      The Brain is a shareware product.  That means you get to try before
      you buy.  If you are satisfied with The Brain and intend to continue
      using it, you are required to register it.  See the 'register.doc'
      file for details on how and where you can register your copy of The
      Brain.

      The Brain v1.2 - User's Manual                                Page 4
      ____________________________________________________________________



				 Program Files
				 -------------

      The distribution package of The Brain contains the following files:

	    - brain12.exe     The stand alone executable version of The
			      Brain.

	    - brain12.doc     The documentation for The Brain (this file).

	    - start-me.bat    A batch file to start The Brain in Beginners
			      Mode.

	    - char123.net     A sample input file to train a network to
			      recognize the numerals 1, 2, and 3.

	    - test123.net     A sample input file to test a network using
			      the numerals 1, 2, and 3.

	    - test123.wts    The weights saved after a network was trained
			     to recognize the numerals 1, 2, and 3.

	    - xor.net        A sample input file to train a network to
			     recognize the 'xor' function.

	    - and.net        A sample input file to train a network to
			     recognize the 'and' function.

	    - vendor.doc     Documentation for vendors, sysops, and others
			     outlining our policy on distributing this
			     program.

	    - license.doc    The license for The Brain.

	    - register.doc   Registration details.  This file explains how to
			     register your copy of The Brain.  Includes
			     addresses for our worldwide registration sites.

	    - order.doc      Order form for The Brain.

	    - support.doc    A list of our worldwide support sites.

	    - nn.faq         The neural network faq (frequently asked questions)
			     from the Internet news group 'comp.ai.neural.nets'.

      The Brain v1.2 - User's Manual                                Page 5
      ____________________________________________________________________


			Introduction to Neural Networks
			===============================


			   What Are Neural Networks?
			   -------------------------

      Expert Definition:

	  A neural network is a parallel distributed information
	  processing structure in the form of a directed graph with the
	  most popular type being feed forward networks.

      Non-Technical Definition:

	  A neural network can be thought of as a pattern recognition
	  system.  The computer learns to associate a certain pattern with
	  a given result.  For example, once a neural network has been
	  taught the characteristics of the numerals 1, 2, and 3, it
	  should be able to recognize those same numerals even if
	  presented in a different font or in a different person's
	  handwriting.


			  How Do Neural Networks Learn?
			  -----------------------------

      Think of a group of interconnected units (as in figure 1).  Data is
      first presented to the system at the input layer in the form of
      zeros and ones (representing 'off' and 'on' respectively).  It then
      goes through a series of hidden mathematical calculations (within
      the hidden layer) before being passed to the output layer, whose job
      is to provide you with some sort of sensible and understandable
      result.

	      |  |
	      O  O Output Layer
	      |/\|
	      O  O Hidden Layer
	      |/\|
	      0  0 Input Layer
	      |  |

	     (Figure 1)

      In general, neural networks can consist of any number of input,
      hidden, and output units, as well as any number of hidden layers.
      However, since any problem can be solved with only one hidden layer,
      and since generalization on unseen data is enhanced with the use of
      only one hidden layer, The Brain restricts you to a single hidden
      layer.  Because of memory considerations, the unregistered version



      The Brain v1.2 - User's Manual                                Page 6
      ____________________________________________________________________

      of The Brain can contain a maximum of 30 units.  The registered
      version allows you to build bigger networks by taking advantage of
      any extra memory you have available.

      Once the information has been processed through the input layer, it
      is fanned out to form the input to each unit in the hidden layer.
      Units in the hidden layer perform a calculation on the input that
      results in a decimal number between 0 and 1. This result then serves
      as the input for the output layer which again performs a calculation
      and produces an output in the range 0 to 1.

      The calculations the network performs in the hidden and output
      layers depend upon decimal numbers known as 'weights'.  Since each
      network performs a unique task, the weights appropriate for that
      network are also unique.

      To determine a network's unique set of weights, the network must
      first be trained to learn how to recognize a specified input and
      output pattern.  For example, to train the network to learn the
      numerals 1, 2, and 3, we must supply the network with a
      representation of those numerals.  We must also tell the network
      what output is appropriate for each numeral.  Thus, if we input the
      number one we would like the network to have the 1st output node
      'on' (i.e.  1) while the other two output nodes are 'off' (i.e.  0).
      When the input is 2, the 2nd output node should be 'on' while the
      other two are 'off', and when the input is 3, the 3rd output node
      should be 'on' while the other two output nodes are 'off'.

      Training a network consists of an iterative process in which the
      network is given the desired inputs along with the correct outputs
      for those inputs.  It then seeks to alter its weights to try and
      produce the correct output (within a reasonable error margin).  If
      it succeeds, it has learned the training set and is ready to perform
      upon previously unseen data.  If it fails to produce the correct
      output it re-reads the input and again tries to produce the correct
      output.  The weights are slightly adjusted during each iteration
      through the training set (known as a training cycle) until the
      appropriate weights have been established.  Depending upon the
      complexity of the task to be learned, many thousands of training
      cycles may be needed for the network to correctly identify the
      training set.

      Once the output is correct the weights can be used with the same
      network on unseen data to examine how well it performs.  For the
      example of learning the numerals 1, 2, and 3, the test data could be
      one of those numerals entered in a different font or in someone
      else's handwriting.

      The whole idea of a neural network is to train the network on an
      input set and then to show the network a similar but different data
      set which it hasn't seen before.  Hopefully the network can



      The Brain v1.2 - User's Manual                                Page 7
      ____________________________________________________________________

      correctly recognize it.  This is very important in areas such as
      handwriting recognition.  While humans can usually read handwriting
      that they've never seen before, this 'simple' task is much more
      difficult for computers.


			    Uses of Neural Networks
			    -----------------------

      The main driving force behind neural network research is the desire
      to create a machine which works similar to the manner our own brain
      works.

      Neural networks have been used in a variety of different areas to
      solve a wide range of problems.  The types of problems solved by (or
      currently being researched using) neural networks include:

	  - voice recognition             - image recognition

	  - stock-market prediction       - car navigation

	  - data compression              - backgammon

	  - character recognition         - chess

	  - horse racing prediction       - sonar recognition

      In theory, neural networks can compute any function a normal
      computer can.  In practice, neural networks are useful for problems
      with a high error rate, that have many examples, and where no
      algorithm exists to solve the problem.

      Professions using neural networks include:

	      - Computer scientists requiring solutions to problems where
		currently no algorithms exist.

	      - Engineers wanting to exploit the capabilities of neural
		networks in their particular application areas.

	      - Cognitive scientists using neural networks to describe
		models of thinking and conscience.

	      - Neuro-physiologists using neural networks to describe and
		explore brain functions.

	      - Physicists using neural networks to model phenomena in
		statistical mechanics.