Artificial Neural Networks and Their Business Appl
Artificial Neural Networks and Their Business Appl
net/publication/222470328
CITATIONS READS
163 273
1 author:
Eldon Li
National Chung Cheng University
119 PUBLICATIONS 3,290 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Eldon Li on 27 July 2021.
by
The Pegasus Group:
Mario Rosso
Phillip Randolph
Appendix A A -0
Introduction and History
World War II and its aftermath produced an explosion of scientific research. During the
War, millions of dollars were spent on developing new technologies in an effort to speed the
process of attaining victory. After the war, however, scientist began to seek ways to build on
One such technology was the computer. While the earliest electronic computers were
little more than calculators, their development was seen as an overwhelming accomplishment in
the field of science and technology. Scientists from around the world rallied around this new
technology, and ideas and theories regarding how it could be used abounded. Of particular
interest to many scientists was the use of computers as a replacement for the human thinking
During the early 1950’s research involving the function of the human brain was
converging with research into computing. While biological scientists were developing theories
on how the brain works, mathematicians were researching the application of these theories to the
new field of computing. In 1956, researchers gathered at the Dartmouth Summer Research
Project on Artificial Intelligence, held at Dartmouth College. The first of its type, this gathering
produced two competing schools of thought for research into the field of Artificial Intelligence.
These were “Expert-Systems” and “Neural Computing.” Neural Computing advocates believed
that the processes by which the brain works could be reproduced through algorithmic modeling
that the process by which artificial intelligence was achieved was not of concern. The real
was able to solicit high levels of funding for research while Neural Computing took a back seat.
Recently, however, neural computing has seen a wide-spread resurgence due to the
limitations of ES. Expert systems typically require a complete understanding of the field in
With ES, an “expert would help to program a computer to mimic that understanding so that it
could ultimately make its own predictions.” Essentially, the expert codifies his human
knowledge into a set of rules which will guide the computer in determining the proper outcomes.
The limitations, however, are that these systems can only be applied where human experts exist.
In areas where there are no real experts, such as in accurately predicting the growth or decline of
the economy, a different approach is required. Many computer scientists believe that neural
computing in the form of Neural Networks may fill this gap in artificial intelligence.
Neural Networks
Professor Kevin Gurney of the University of Sheffield, defines a Neural Network as “an
loosely based on the animal neuron. The processing ability of the network is stored in the inter-
unit connection strengths, or weights, obtained by a process of adaptation to, or learning from, a
Neural Networks, in contrast to Expert systems, do not require an explicit set of rules.
Instead, the logic behind Neural Networks stems from the accepted biological model of how the
brain works. The network makes up its own rules that match the data it receives to the result it's
told is correct.
Neural Networks, like the brain, consist of numerous artificial “neurons” or processor
units. Each unit receives inputs, both excitatory and inhibitory, from a number of other units
and, if the strength of the signal exceeds a given threshold, sends signals to other units. The
action, so to speak, is in the connections among these units. Herb Brody of Technology Review
states, “Each neuron takes many input values, multiplies each by a "weighting" factor, adds
together the products of these multiplications, and mathematically operates on this sum to
produce another set of values. These numbers may translate directly into an answer--say, the
The term “Neural Network” is inherently a biological one. While the term Neural
Network can be used to define either a network of interconnected, organic nerve cells in
biological organisms, the subject of this work is to identify attributes and applications of
computer-based Neural Network systems. A more correct term for software-based Neural
Networks would be Artificial Neural Networks or ANN. For the purpose of simplicity, from this
point forward, the term Neural Network is used throughout this work to describe computer-based
ANNs.
Initially, a Neural Network must undergo a period of training to develop its “own rules.”
During the training phase, the network is fed a stream or batch of input data, and it determines
the strength or weight of the relationships between each variable and a given output. The
training process determines input weights that will allow a Neural Network to function properly
based on the training data. Provided that you train a Neural Network with appropriate and
complete data, the network will determine appropriate outputs from inputs that are slightly
different from inputs that are entered during training. In other words, as the patterns of
relationships are determined between the inputs and the given outputs, the network can predict
The real “intelligence” behind Neural Networks is its ability to automatically make
corrections in the weights and relationships it gives to each variable entered. This is through a
interaction effects, and weights of each variable as they are related to a given output, a set of
“rules” is developed. However, if more data is entered into the system in the form of inputs and
outputs, the network will test its set of rules. If the rules do not produced the desired outputs, the
network goes back to determine what rules need to be changed. In essence, the backpropagation
process adjusts the weights of the various relationships until they meet the given output. The
weight of each relationship or variable is increased or decreased in small increments until the
weights that produce the proper output are determined. As the proper output is attained, the
network develops a new set of “rules” to guide itself in making new outputs.
Data Mining
Dramatic reductions in the cost of data storage have facilitated a substantial increase in
the amount of data currently being collected and stored by organizations. In order for managers
to gain value from this stored data and to improve the decision-making process, hidden patterns
must be discovered. Neural networks are well-suited to the task of recognizing patterns in
enormous amounts of data such as is commonly found in data warehouses. Some of the
attributes of Neural Networks that make possible effective data mining activities include the
ability to explore all possible interrelationships among key elements, model complicated
problems and multi-dimensional data and execute pattern extraction with greater speed than
conventional tools.
Marketing
planning. Issues such as ordering of raw materials, manpower requirements and inventory levels
depend on accurate and timely retail sales forecasts. Traditionally, Multiple Regression Analysis
(MRA) was performed to forecast sales. Utilizing Neural Network methods instead of MRA
Independent variables present at the input layer of a retail sales forecasting Neural
Network might include market demand for product, disposable income of consumers, size of
population, price of product, price of substitute products and price of complementary products.
Dependent variables at the output layer will include sales and the amount of time to reach a
break-even point.
Within the hidden layers of the Neural Network, weights for the network can be
initialized to arbitrary values. During the training phase, historical data is cycled through the
network a number of times. As differences are identified between computed outcomes and
correct historical data, the weights within the hidden layers will change and adapt. Finally,
actual data can be fed into the network and reliable forecasts can be computed.
Direct marketing is another application better served by Neural Network techniques than
traditional statistical methods. One of the objectives of direct marketing is to minimize the
amount of mailings to people who are not as likely to purchase the goods and services being
Discriminant Analysis; however, Neural Network has proven to be a more efficient method for
identifying prospective customers. Independent variables such as age, sex, income, occupation,
education, social class and geographic location serve as the inputs for the network. The outputs
Target marketing seeks to segment a large market into one or more smaller groups. After
division, marketing campaigns can be focused on each segment independently . Cluster Analysis
is traditionally used as the conventional tool for market segmentation. In applying Neural
toward product are presented at the input layer. The outputs of the network will include the
Manufacturing
The pattern recognition capabilities of Neural Networks make them well-suited for
optimization and failure prediction are becoming increasingly common and accepted.
Lucent Technologies, Inc. combines Neural Network with genetic algorithms to improve
process control in its semiconductor fabrication labs. Variations in film thickness increase
production costs and reduce semiconductor yields. The complex, 450-step production process
involved in the fabrication of Lucent’s semiconductors, combined with the small tolerances of
the devices, make it nearly impossible to identify which minute adjustments would have the most
impact on improving process control. To solve this problem, engineers trained a Neural Network
with plasma etching time data taken from a database of previous production lots. The resulting
petroleum industry. Quality attributes for gasoline, such as octane, can be inferred through the
use of Neural Networks. Traditionally, such measures of quality had to be obtained during
processing using physical analyzers such as gas chromatographs. This physical analysis method
caused delays in production and required the use of expensive, maintenance- intensive
equipment. Neural networks utilizing simple inputs such as temperature, flow rates and pressure
have reduced, and in some cases eliminated, the need for online physical analysis methods.
setting of warranty policies. By loading a Neural Network with sensor data obtained in new car
stress testing and then comparing this data with the patterns of previous model tests, durability
information can be deduced The resulting mean time between failure (MTBF) calculations are
Development System (NDS) software package. Designed by Nestor, Inc. to run on personal
computers and Sun Microsystems workstations, it has the ability to recognize visual patterns for
part identification. The system can recognize parts even after slight modifications have been
made in their shape or in changing lighting environments. The NDS package also has
signals.
Financial
Forecasting and prediction are important elements of investment decision making. While
most investment decisions are made by individuals with expertise in the field of finance, Neural
Networks are beginning to challenge the performance of human fund managers. Of the many
factors which make Neural Networks well suited to forecasting applications, the ability to
analyze enormous amounts of historical performance data and to identify relationships and
patterns in that data are of paramount importance. Data associated with financial markets is
readily available and can easily be codified for use in computerized Neural Network systems. In
addition, human factors such as ego and caution that may influence a decision maker’s judgment
Financial forecasting Neural Networks can be separated into two categories. The first
group is concerned with prediction of future events from the analysis of known facts. Examples
of this type of Neural Network include programs that are designed to predict bankruptcies or
potential loan defaults. A second category of forecasting Neural Network is concerned with
Time Series Forecasting. Time Series Forecasting examines past events and data and attempts to
predict future trends. An example of this type of Neural Network application includes systems
that are used to predict stock and bond market trends and prices.
As is the case with all Neural Networks, an investment forecasting Neural Network must
be trained to recognize positive and negative relationships to produce useful outputs. Training is
learning involves comparing the outputs of the Neural Network to actual historical outputs and
adjusting the weights of the hidden layer to minimize the difference between actual and
calculated results. Unsupervised learning does not involve the comparison of calculated and
actual outputs. Instead, unsupervised learning methods rely on the network itself to find positive
relationships within the large amounts of data that are fed into it.
examples are provided herein. The first example involves an experiment conducted in 1990
involving stock market prediction. Using a Neural Network to analyze data from the Tokyo
Exchange, researchers were able to predict the future direction pattern of the market with 94
percent accuracy. A second example involves bond rating. A sample of bond issues from 40
companies were used to test a Neural Network’s ability to rate bonds. Thirty of the bond issues
were used as input data to train the network while the remaining ten were used to compare
against the results generated by the network. The Neural Network was able to predict the bond
ratings with an accuracy rate of 88 percent, compared with only 65 percent for regression
analysis techniques. Third in the list of examples involves an Neural Network used by LBS
Capital Management Inc. in Clearwater, Florida. The system was used to pick stocks for
purchasing between the years 1991 and 1993. It performed consistently better than the Standard
& Poor’s 400 Index each year. A final example involves the commodities trading market. A
Neural Network trained with three years of trading data realized a 52 percent profit after
operating for six months on an original investment of $3,000. Figure 1 below illustrates a
90%
80%
70% S&P 500 Euro
60% NIKKEI Crude Oil Yen
50%
Futures
40%
30%
20%
10%
0%
Applications in risk analysis stand out as some of the most successful and wide-spread
Neural Network business applications developed to date. Many large financial institutions ha ve
incorporated Neural Network based fraud detection systems into their organizations. One such
system, Visa U.S.A.’s Cardholder Risk Identification Service, produces an alert when irregular
Credit ratings can also be ge nerated using systems such as Neural Technologies’ Decider.
Standard & Poors (S&P) is currently using Decider as the engine powering its Neural Network
based CreditModel system. CreditModel is an Internet delivered service that allows financial
institutions doing business with S&P to rate the credit risk of potential customers.
CreditModel’s inputs include the customer’s revenues, debt load, and prior S&P credit ratings.
Using CreditModel’s Neural Network to determine a bank’s credit rating has brought the cost of
S&P’s service down from $40,000, using human analysis, to $400. Similarly, the process of
Sears Mortgage Corporation is using Neural Networks to streamline and augment its
by its Neural Network, while human underwriters scrutinize the remaining “tough” applications.
Among the thirty pieces of data input into the Sears’ neural net are applicant’s zip code, age of
the home, appraised value of the home, monthly income, proposed housing expense and credit
rating. The network is initially trained by running iterations on 2,500 application acceptances
and 2,500 rejections. Outputs simply establish whether the application is to be accepted or
rejected.
Other Applications
The number of fields benefiting from Neural Network technology is vast, and identifying
them all is beyond the scope of this report. Many of these applications are difficult to classify
but deserve to be noted. Among these are systems that utilize the pattern recognition strengths of
Neural Networks in applications such as the reading of Japanese Kanji characters and English
hand writing on checks, identifying battle-field targets and classifying radar signals, and locating
features on seismic maps that aid petrochemical industry geologists in the exploration of new
natural resources.
Two of the most surprising and interesting applications of Neural Networks technology
discovered while researching this report include race horse paternity determination and network
combining Neural Network powered pattern recognition systems with video cameras to verify
race horse pedigree. Proof of parentage, verified through a series of 142 reaction tests on a
Davis were becoming overwhelmed by as many as 72,000 reaction tests in a single day. To
solve this problem, scientists at Davis installed a Neural Network system produced by San
Neural network based failure prediction and alert systems, such as Computer Associates’
Neugents product, seek to alert engineers of impending failures before an event actually occurs.
Neugents software is currently being used to identify and learn historical patterns in data network
infrastructures to facilitate preventative corrective action before a system fails or its performance
degrades.
field with enormous potential. While still largely unknown or misunderstood, this rapidly
evolving field holds promise for a vast number of tasks that were previously accomplished
through human cognitive activity. Throughout the course of this study, Neural Networks were
defined and a number applications were highlighted. While not an exhaustive list, many of these
applications, such as those in fraud detection, process control, market prediction and pattern
recognition, should give the reader an introductory understanding of Artificial Neural Networks
and their amazing potential. Neural Networks are at last being touted as a worthy alternative to
their Artificial Intelligence counterparts, Expert Systems. In the future, humans will delegate
more and more of their cognitive and decision making activities to Neural Networks and Expert
Systems. This fact should make for some exciting possibilities and concerns that deserve further
study.
Bibliography
Brody, Herb. “The Neural Computer.” Technology Review 93, no. 6 (August-Sept 1990):
42-49.
Cheng, Wei, Bruce W. McClain and Christopher Kelly. “Artificial Neural Networks
Make Their Mark as a Powerful Tool for Investors.” Review of Business 18, no. 4
(Summer 1997): 4-10.
Davidson, Clive. “A Mind of their Own.” Computer Weekly, 14 Jan 1993, 26-24.
Enticknap, Nicholas. “Knowledge is the Key.” Computer Weekly, 30 Nov 1995, 48-49.
“Fab Control gets Neural Guidance.” Electronic Engineering Times, 16 December 1996.
Getz, Arlene. “Out in The Cold: Travel nightmare: far from home with a frozen credit-
card account.” Newsweek, 30 Oct. 2000, 102.
Gianturco, Michael. “The Eye is a Computer.” Forbes 146, no. 1. 12 Nov 1990, 32-35.
Hoffman, Thomas. “Neural Nets Spot Credit Risks.” Computerworld, 16 July 1999, 38.
Knight, Kevin. “Connectionist Ideas and Algorithms.” Communications of the ACM 33,
no. 11 (Nov 1990): 58-75.
Quitttner, Jeremy. “Credit Union Unit Gets New Tools to Fight Fraud.” American
Banker 161, no. 64, 3 April 1996, 15.
Roberts, Leslie. “Are Neural Nets Like the Human Brain?” Science 243, no 4890 (Jan
27, 1989): 481-482.
Romberg, Denise. “CA’s ‘Neugents’ Forcast Failures.” Computing Canada 25, 1 Jan
1999, 17.
Smith, J. Clarke. “A Neural Network – Could it Work for You?” Financial Executive 6,
no. 3 (May-June 1990): 26-31.
Wright, Maury. “Neural Networks Tackle Real-World Problems.” EDN, 8 Nov 1990, 79-
87.
APPENDIX A Commercial Neural Network Software Applications
(source: http://www.emsl.pnl.gov:2080/proj/neuron/neural/products/)
Marketing
Real Estate
Market Trading
Fraud Detection
Process Controllers
• Fujitsu
o Application: controller for continuous casting
• Honeywell
o Application: process control systems
• Neural Applications Corporation
o Application: controller for steel making
o Product: Intelligent Arc Furnace
• Nippon Steel
o Application: controller for continuous casting
• Pavilion Technologies
o Application: process control, waste reduction
o Product: Process Insights®
o Customers: Eastman Kodak
• Siemens
o Application: process control of rolling mills
o Customers: Krupp Hoesch Stahl, HYLSA, Thyssen, Voest Alpine, Hanbo,
Gallatin Steel,
• Texaco's Puget Sound Refinery
o Application: refinery process control system
• Anheuser-Busch
o Application: beer testing
• Applied Intelligent Systems
o Application: vision recognition systems
• CTS Electronics
o Application: loudspeaker defect classifications
• Dunlop
o Application: tire testing
• Intel
o Application: computer chip manufacturing quality control
• Monsanto
o Application: predicting quality of plastics
• Netrologic, Inc.
o Application: welding quality control
• NLK - Celpap
o Application: predicting quality of paper
• Volvo
o Application: diesel knock testing, paint inspection