From tr@fct.unl.pt Tue Dec  7 01:34:23 EST 1993
Article: 19778 of comp.ai
Xref: glinda.oz.cs.cmu.edu comp.ai.neural-nets:13765 comp.ai:19778
Newsgroups: comp.ai.neural-nets,comp.ai
Path: honeydew.srv.cs.cmu.edu!rochester!udel!darwin.sura.net!howland.reston.ans.net!pipex!uknet!EU.net!news.inesc.pt!dec4pt.puug.pt!unl!news
From: tr@fct.unl.pt (Thomas Rauber)
Subject: FEATURE SELECTION SOFTWARE AVAILABLE
Message-ID: <1993Dec3.125954.1157@fct.unl.pt>
Keywords: Feature selection, supervised classification
Sender: news@fct.unl.pt (USENET News System)
Organization: Universidade Nova de Lisboa, PORTUGAL
Date: Fri, 3 Dec 1993 12:59:54 GMT
Lines: 91

----------------------------------------------------------
*****                                                 ****
*****      FEATURE SELECTION SOFTWARE AVAILABLE       ****
*****                                                 ****
*****                tooldiag 1.4                     ****
*****                                                 ****
----------------------------------------------------------

MOTIVATION:
----------

Classifiers that use ALL possible features are in general
 - more complex
 - less reliable
than classifiers that use only a subset of all possible features.
For instance a neural network that wants to classify a 10-class problem
using its 512-valued Fourier spectrum and a F---(2F+1)---C
fully interconnected 3-layer feedforward architecture, has 1546 neurons,
where F is the number of features and C is the number of classes.

If only the 10 most relevant features were selected, the complexity
of the net would reduce to 41 neurons. Furthermore its classification
accuracy would eventually be increased.


FEATURE SELECTION:
-----------------

The software package "tooldiag" performs a feature selection. Many concepts
of the book:
  Devijver, P. A., and Kittler, J., "Pattern Recognition --- A Statistical
  Approach," Prentice/Hall Int., London, 1982.
are implemented, including the optimal BRANCH & BOUND search strategy,
together with several different selection criteria.


ADDITIONAL CAPABILITIES:
------------------------
 - An error estimation can be performed, using the Leave-One-Out method
   and a K-Nearest-Neighbor classifier.
 - A learning module (Q*) is included that has the same functionality
   as the LVQ (Learning Vector Quantization)


INTERFACING:
------------

The system has interfaces to

	- LVQ_PAK - The implementation of the Learning Vector Quantization
			(see FAQ of comp.ai.neural-nets)
	- SNNS - The Stuttgart Neural Network Simulator
		A pattern file (.pat) can be generated, using only the selected
		subset of features. Besides a simple F---(2F+1)---C backprop
		net is generated (.net).

The data file format is compatible with that of LVQ_PAK.
2-D graphics are displayed with the help of the GNUPLOT public domain
plotting package.


RESTRICTIONS:
-------------
1.) Only continuous (or ordered discrete) numerical features
2.) No missing values


HOW TO GET IT:
-------------

A documentation and the source code in C is provided. The system was tested
on many platforms (IBM, DEC, NeXT, SUN - workstations, DOS ) and is very
easy to install.

Location: Anonymous FTP at:
SERVER: ftp.fct.unl.pt
DIRECTORY:  pub/di/packages
FILE:  tooldiag-1.4.tar.Z


Enjoy

--
 ________________________________________________________________________
|                                  |                                     |
|         Thomas W. Rauber         | BITNET/Internet: tr@fct.unl.pt      |
|__________________________________|                                     |
|   Universidade Nova de Lisboa    | Fax:   (+351) (1) 295-7786          |
|   Intelligent Robotics Center    | Phone: (+351) (1) 295-7787          |
|  2825 Monte Caparica, PORTUGAL   |                                     |
|__________________________________|_____________________________________|