[2091] in comp.os.os2.announce archive
VENDOR: Software Update -- NeuralCrunch/2 v2.08.02 (Neural Net Modelling)
daemon@ATHENA.MIT.EDU (Dennis R. Grace)
Sat Nov 16 16:58:07 1996
To: os2ann.DISCUSS@bloom-picayune.MIT.EDU
Date: 13 Nov 1996 17:33:22 GMT
From: dgrace@gpn.com (Dennis R. Grace)
Reply-To: dgrace@gpn.com (Dennis R. Grace)
Reply-to: dgrace@gpn.com (Dennis R. Grace)
---------------------------------------------------------------------
Subject: VENDOR: Software Update -- NeuralCrunch/2 v2.08.02
Grace Pacific Numerics is happy to announce the release of version 2.08.02 of
NeuralCrunch/2, the multithreaded, 32-bit neural network development utility
for OS/2. NeuralCrunch/2 is designed to allow a general user to leap right
into using and exploring neural network technology. No prior experience with
neural networks is assumed, but the experienced user should find both the
online and reference documentation to be fairly complete in their discussion
of the relevant mathematical issues involved in initialization and training.
Context sensitive help is available online for every configurable feature.
Artificial neural networks have generated great interest among scientists,
engineers and financial modelers due to their demonstrated ability to provide
insight into complex systems. Multilayer feedforward networks especially have
proven themselves in approximation and prediction problems. It is the purpose
of NeuralCrunch/2 to provide both the novice and expert direct access to this
technology.
"For those interested in neural network experimentation and a level of
complex-system modeling not available with any other type of tool,
NeuralCrunch/2 is a straightforward and powerful product."
--- Brian Proffit
Senior Contributing Editor
OS/2 Magazine
October 1996
COMPLIMENTARY UPGRADE TO CELEBRATE MERLIN
NeuralCrunch/2 v2.08.02 represents a complete refresh and optimization for
OS/2 Warp v4. To mark the release of OS/2 Warp v4 (aka Merlin), complimentary
upgrades to NeuralCrunch/2 v2.08.02 have been shipped to all registered users
of earlier versions; if you purchased a copy of an earlier version and did not
bother to register, now would be a good time to do so.
AVAILABILITY
NeuralCrunch/2 v2.08.02 is now available directly from Indelible Blue, whom
you may reach at (800) 776-8284, or via their web catalog at
http://www.indelible-blue.com/ib. NeuralCrunch/2 has a suggested retail price
of US$49.95.
Demonstration versions are currently available at the Walnut Creek
Archive (ftp://wcarchive.cdrom.com/pub/os2/demos/ncrunch2.zip)
and HOBBES (ftp://hobbes.nmsu.edu/os2/demos/ncrunch2.zip).
FEATURES
The value of NeuralCrunch/2 is revealed in its generality and
complete configurability.
* Simple numerical text files can be used as sources of data for training
and evaluation. NeuralCrunch/2 allows both data and network coefficients
to be saved and reloaded independently.
* There is no essential restriction on the size of the data sets and the
number of hidden nodes used in NeuralCrunch/2.
* NeuralCrunch/2 uses multilayer feedforward networks with a single hidden
layer and logistic squashing functions in the hidden nodes. This class of
neural network has been rigorously shown to be capable of approximating
arbitrary functions and their derivatives arbitrarily well. A growing body
of practical experience substantiates this theoretical capability.
* A number of network initialization methodologies are provided, including
the use of principal components, a very common, more conventional method
of data analysis.
* The network topology is dynamically configurable, meaning that any number
of additional inputs, outputs and hidden nodes may be formed at any point,
and any number of current inputs, outputs and nodes may be removed at any
point.
* All network data and coefficients are presented in completely configurable
graphical views, as are comparisons of desired and actual network outputs,
node responses and output dependencies. Most graphical views can be
enlarged and scrolled. The multithreaded architecture of OS/2 allows
computationally intensive activities such as training to be separated from
both the user interface and the graphical presentation of this
information. This allows the various features of a network to be
investigated while training is actually in progress. All graphical views
can be printed.
* Among the most powerful, globally convergent methods for minimizing
nonlinear systems which are currently available, a modified Newton-Raphson
algorithm, with line-searching and backtracking, is used to evaluate the
network coefficients as functions of the training data. This algorithm is
numerically stable, tractable, efficient and general. Annealing may also
be employed in training.
* The performance of the network is monitored in a number of ways, all of
which can be used to control training.
The interface of NeuralCrunch/2 includes a standard action or menu bar, a
status line at the top of the viewing area, and a number of the more
frequently invoked actions are also represented in a collection of pushbutton
icons located along the bottom of the viewing area.
The authors of NeuralCrunch/2 think you will find this utility a valuable
addition to all of your spreadsheet, financial analysis, statistical, signal
processing and other data analysis software.
SYSTEM REQUIREMENTS
NeuralCrunch/2 is a true 32-bit, multithreaded, native OS/2 application which
will run under IBM OS/2 versions 2.1 and above, especially Warp v3/v4, on at
least i486DX compatible systems with a minimum of eight megabytes of memory,
at least a generic VGA adapter, and something less than three megabytes of
disk space. As with any complex piece of software being used under a
multitasking operating system, the more memory the system has the better.
Likewise, NeuralCrunch/2 is optimized for performance on Warp/Pentium-based
systems, so you will be happiest using Warp and at least some sort of Pentium.
There is no essential restriction on the size of the data sets and networks
which can be used in NeuralCrunch/2, but for larger problems you should ensure
that your swap file has plenty of room to grow. Some sort of printer is
required for hardcopy output.
Grace Pacific Numerics
3137 Fenelon Street San Diego, CA 92106-2209 United States of America
NeuralCrunch/2 version 2.08.02 (October 1996)
Copyright Grace Pacific Numerics 1996.
All rights reserved.
=====================================================================
-- To submit to c.o.os2.announce: POST (or email os2_ann_req@bix.com)
-- For comments to the moderator: email lfirrantello@bix.com.
- for posting guidelines see: http://www.bix.com/pub/os2ann/pindex.htm
