Printer Friendly

AI-based neural networks.

Al-Based Neural Networks

The latest in information-processing technology

Just as expert systems sparked interest when they first emerged into the business computing marketplace, neural networks are now in the information technology limelight. Current market forecasts seem to indicate that neural networks will be a significant part of the future investment in information technology. The neural network market, which amounted to about $7 million in 1987, is forecast to reach $365 million in 1993.

Much can be learned about these networks by looking at some prototype neural network implementations that have been developed in the area of business applications. It is important to underscore the term "prototype implementations," because there are very few full-scale marketable systems that have been developed in this new phase of information-processing technology.


Neural networks are essentially a type of information-processing technology. These systems take in information, process it, and produce output similar to other forms of business computing applications. As depicted in the illustration on this page, what makes neural network systems different is that their design is inspired by studies of the brain and nervous system.

Neural networks are made up of many simple, highly-interconnected processing elements that dynamically interact with each other to "learn" or "respond to" information rather than to simply carry out algorithmic steps or programmed instructions. Information is represented in a neural network in the pattern of interconnection strengths among the processing elements. Information is processed by a changing pattern of activity distributed across many units. Learning occurs through an interactive adjustment of interconnection strengths based upon information within a learning sample.

The illustration on this page shows the previously mentioned features of neural networks and shows how an array or sequence of numbers is entered into the network to produce an output array. As shown in this multilayered network, each processing element in the first layer takes a component of the input sequence, operates on it in parallel with the other processing elements in the layer, and delivers a single output to processing elements in a layer above. The result is an output sequence that usually represents some specific characteristic associated with the input.


Neural networks are best at classifying patterns based upon previous experience and learning complex relationships among elements in an information base. Pattern classification, for example, enables a neural network system to make a "judgment" about the risks involved in extending credit to a particular client or of investing in a certain portfolio of bonds.

The capability of neural networks to synthesize continuous-valued data allows neural networks to "discover" relationships between several inputs and one or more outputs based upon some number of sample data points. Once such relationships have been found and represented within the network, this "knowledge" is used to interpolate between or extrapolate beyond the sample data. Applications include filtering noise out of EKG signals and creating economic forecasts for such things as interest rates, stock prices, and consumer production. An examination of the following implementations may help to illustrate the versatility of neural networks.


The mortgage insurance industry is considering ways of using artificial intelligence technology to assist in the process of mortgage risk assessment. Poor judgments made by mortgage underwriters usually lead to major financial losses. The development of rule-based expert systems for mortgage insurance applications has had less success than for the general population of mortgage applicants. The reason is that in the mortgage insurance cases (the higher risk group), the risk assessment rules used by underwriting "experts" tend to be more subtle and are plagued with more ambiguities. The high economic value of reducing the number of bad risks has led some in the mortgage insurance industry to begin to develop neural network decision systems.

Nestor, Inc., of Providence, Rhode Island, claims to be one of the first to offer neural network products to appraise mortgage applications. The advantage of neural networks over expert systems was stated by a Nestor spokesman: "The beauty of neural networks is you don't have to write the rules--it learns from the data."

Nestor has developed and tested what they call a Multiple Neural Network Learning System (MNNLS). A prototype implementation has been trained using human underwriter judgments. The system learned to mimic underwriting by capturing the complex relationships in the data. The training data set consisted of 5,048 applications from all parts of the U.S. during the period September 1987 to December 1987. About half of these applications were accepted for mortgage insurance. Attributes in the training data set included information on the borrower's "cultural" status (e.g., credit rating, years employed, and number of dependents) and financial status (e.g., income and obligations). The data set also included information about the mortgage itself (e.g., loan-to-value ratio, ratio of income-to-mortgage payment and loan amount), as well as information about the property (e.g., age, location, and appraised value).

When the MNNLS was tested on a data set of previously unseen examples, the level of agreement with human underwriters ranged from 82% to 96% depending upon the degree of consensus required between the system's various levels of subnetworks in the multiple network learning system. Just as in a multiple-level human decision making organization, requiring a higher degree of consensus results in a smaller percentage of the cases being unambiguously identified, but those that are identified have a higher level of accuracy.


Forecasting is a fundamental task of business decision making. Future sales is frequently the most important variable in business forecasts. Economic forecasts for such other quantities as interest rates, stock prices, exchange rates, and so forth also play a critical role in the planning and management of businesses. Many statistical and operations research methods have been used as forecasting techniques, which often fail when they are unable to capture the complex relationship in the historical, time-series data. Neural networks can be applied to these situations.

The chart to the right shows a simplified example in which a system developed by Neuralware, Inc., of Pittsburgh was used to predict stock prices. This example is based upon the week's-end closing price of Standard and Poor's S&P 500. The network was trained to predict the closing price for the following week, given the closing prices for the prior ten weeks. The lower curve in the figure shows the actual data. The other two curves illustrate a comparison between the neural network forecasting model (middle curve) and a ten-week moving average forecasting procedure (upper curve). The ability to correctly predict the trend serves as a basis for comparison between the two approaches. In this example, the neural network predicted the trend (up or down) 61% of the time compared to 53% for the ten-week moving average procedure.


The ability to determine a bond rating that accurately reflects the risk of investment in that particular bond is of critical importance to bond issuers for two reasons. First, the bond rating directly affects investors' bond purchase decisions. Bank investments, for example, are restricted to top-rated bonds. Second, the rating has a significant effect on the offering yield of the bond itself.

Considerable effort has been expended by rating agencies in developing models that will reliably perform the required rating of bonds. A common method has been to use statistical techniques dealing with such financial ratio variables as net worth over total debt, sales over net worth, profit over sales, debt over total capital, income over assets, and income over interest charge. These models have yielded relatively poor results. The use of expert systems has had even less success because of the difficulty of capturing and representing the knowledge of experts in this highly competitive and sensitive field.

Researchers at the University of California in Berkeley have applied a neural network model based upon ten financial variables and have compared results with statistical techniques using the same data. The neural network model was able to correctly classify over 80% of the bonds correctly, while the statistical model had an accuracy rate of only about 60%.

There are a surprising number of encouraging results in the application of this new information technology in business computing applications. Unfortunately, very little detail is made available on business-oriented neural network implementations because of the substantial economic potential and the high degree of confidentiality of much of the information in these areas.

This lack of detail may force managers to explore for themselves the potential usefulness of neural networks rather than relying upon the experience of others. However, the availability of artificial neural network simulators that run on conventional computers makes the cost for developing simple prototype implementations relatively small.

There is indeed a lot of hype surrounding neural networks, but results of early implementation in the field of business computing suggest that business managers should view this latest advance in business computing technology with cautious optimism.

PHOTO : Biological neurons with dendrites leading into an artificial neural network. Synapses

PHOTO : connect biological neurons. Artificial neurons (the empty circles) are held together

PHOTO : with connection weights (the smaller black circles).

Dr. Pracht is associate professor of Management Information Systems and Decision Sciences at Memphis State University.
COPYRIGHT 1990 University of Memphis
No portion of this article can be reproduced without the express written permission from the copyright holder.
Copyright 1990 Gale, Cengage Learning. All rights reserved.

Article Details
Printer friendly Cite/link Email Feedback
Author:Pracht, William E.
Publication:Business Perspectives
Date:Mar 22, 1990
Previous Article:Telecommunication systems of tomorrow.
Next Article:Model management systems in business.

Related Articles
Neural networks: the buck stops here.
High society on the brain.
The brain in the machine: biologically inspired computer models renew debates over the nature of thought.
The AI factory; how artificial intelligence will create 'smart plants.' (Cover Story)
Neural-net neighbors learn from each other.
Automatic flat dies gain artificial intelligence.
A computer eyes the heavens.
Artificial Neural Networks make their mark as a powerful tool for investors.
Artificial intelligence in accounting and business.
Neural network forecasting of the production level of Chinese construction industry.

Terms of use | Copyright © 2016 Farlex, Inc. | Feedback | For webmasters