Advanced Neural Network and Genetic Algorithm Software

Prediction of Long Term Responses to HIV Infections from Short Term Measurements Using Neural Networks
In the Frederick County Maryland 2002 Science Fair, high school student Jennifer Gee won the grand prize and a trip to the 53rd International Science Fair with this project. The project also won first place in the Computer Science category.

"You'll be happy to hear that I used your AI Trilogy to make predictions on the efficiency of HIV drug treatments", Jennifer told us. "This is the same project that I worked on in the summer with Dr. Dimitrov." Dr. Dimitrov is a leading researcher at the NIH Frederick Cancer Research Center.

Jennifer was also awarded an Intel Excellence in Computer Science Award from Intel, a U.S. Army Achievement Award, a U.S. Navy and Marine Corps award, a US Dept of Health Award, and also a Yale Science and Engineering award.

In May 2002 Jennifer went on to compete in the international science fair in Louisville, Kentucky. There she won two prizes from the special awards sponsors. She received a $1000 award from the Association of Artificial Intelligence, and $200 from the Association of Computing Machinery.

Researcher uses WSG neural networks in the study and treatment of rheumatoid arthritis (RA).
Long time user Guirish Solanki is a researcher at the Dept of Surgical Neurology, The National Hospital, Queen Square, London. Since the mid-nineties he and his colleagues have published several impressive papers describing use of our neural networks in the study and treatment of rheumatoid arthritis (RA). Click here to view.

SUPERIORITY OF NEURAL NETWORKS OVER DISKRIMINANT ANALYSIS METHODS FOR HISTOLOGICAL CLASSIFICATION OF HUMAN MENINGIOMAS
José Rafael Iglesias-Rozas (1), Frieder Albert Grieshaber (2) Alexander Sobol (2) (1) Pathologisches Institut. Katharinenhospital (Ärztl. Direkt. Prof. Dr. med. B. Kraus-Huonder =). Kriegsbergstrasse 60, D-70174 Stuttgart.(2) Institut fuer Biomedizinische Technik (Direkt. Prof. Dr.Ing.U. Faust), Seidentrasse 36, D-70174 Stuttgart. (Electronic Journal of Pathology and Histology (1996) Volume 2/1:961-05.txt (1-9)

ABSTRACT
In the present study, the utilities of Backpropagation, Probabilistic and General Regression neural networks (NNs) were compared with discriminant analysis methods in classifying human meningiomas based on histological characteristics. 813 cases of human meningiomas were selected from 5200 human brain tumors and histologically evaluated. In all cases the presence and the subjective quantity (0, 1, 2, 3) of 50 histological and 15 clinical characteristics were observed by a competent neuropathologist. The cases were sorted into ten groups according to the international WHO-classification: endotheliomatous, fibroblastic, transitional, psammomatous, angiomatous, angioblastic, papillary, malignant, meningosarcoma and "others". Artificial NNs were created using development software (Neuro Shell 2.1) on Windows TM .To compare the results with NNs, 217 testing meningiomas were classified based on the principle of discriminant analysis (DA). The primary classification of testing patterns of meningiomas into ten original groups in our series produced very insufficient results with both NNs and DA. The stepwise classification with BPNNs are obtained using 'a hierarchic model of the best classified pattern group': Diagnoses with very heavy variables were separated from the rest of diagnoses to build a separate group. From the rest of other diagnoses by using the module 'contribution factors/variables' on NS21 the heavier variables were searched to obtain the best tumour classification, and so on. The module 'contribution factors/variables' shows 'a rough' measure of the importance of that variable in predicting the networks output. At the end of this stepwise principle, all meningiomas were classified with an average efficiency of 82.5% and 73.6% with BPNNs and DA respectively. Classification methods based on NNs are practical and relatively inexpensive. If trained NNs are developed and validated, classification can be performed with minimal data entry skills by different institutes and neuropathologists. An international classification and evaluation of human tumours can thus be done more effectively. NNs techniques appear promising in the field of histological pathology, particularly in combination with complex laboratory and clinical data, should be investigated with more emphasis. (Electronic Journal of Pathology and Histology (1996) Volume 2/1:961-05.txt (1-9)

KEYWORDS: Meningiomas, Classification, Neural Networks, Discriminant Analysis.

Diagnosis of Human Oligodendrogliomas with the Help of the NeuroShell Easy ClassifierTM Neural Network
Authors: José Iglesias, M.D., Ph.D., Javier Esparza, Prof. Inf., and Mathias Scherf, Dr. Inf.

Objective: To examine whether a suitable solution can be found concerning the ability to reproduce the histologic classification of human oligodendrogliomas with the assistance of the NeuroShell Easy ClassifierTM neural network.

Study Design: Histologic sections of 449 human oligodendrogliomas were selected. The diagnostic task was given by differentiation of three oligodendroglioma types: 121 low grade oligodendrogliomas, World Health Organization grade 2; 180 low grade oligoastrocytomas; and 148 anaplastic oligodendrogliomas, grade 3. Age, sex and 50 histologic characteristics were examined in each case, describing the presence of a specific histologic feature on a scale of four (zero, absence of the feature; three, abundant presence). From each group, two-thirds of randomly selected tumors were available for the training set and one-third for the testing set.

Results: In the three-class problem, 98.88% of the tumors were correctly classified (testing set). Ninety-nine percent of new testing tumors were correctly classified with Easy ClassifierTM as low grade and anaplastic oligodendrogliomas. In the case of low grade oligodendrogliomas versus low grade oligoastrocytomas, 99% of new tumors were correctly classified.

Conclusion: The main conclusion from this study is that Easy ClassifierTM was able to differentiate, with high accuracy, sensitivity and specificity, among the three types of oligodendrogliomas. (Analyt Quant Cytol Histol 2000;22:383-392)

Keywords:oligodendrogliomas, neural networks (computers), NeuroShell Easy ClassifierTM

Priv.-Doz. Dr. J.R. Iglesias-Rozas
Klinikum Stuttgart, Katharinenhospital
Institut für Pathologie. Neuropathologie
(Ärztl. Leiter: Prof. Dr. med. A Bosse)
Kriegsbergstr. 60
D - 70174 Stuttgart
Tel: + 711 278 4918
Fax.: + 711 278 4909
e-mail: jr.iglesias@katharinenhospital.de

Here is a more complete list of Dr. Iglesias' publications using NeuroShell:

1. BRODBECK; A.; IGLESIAS-ROZAS, J.R. ; ZELL, A.:
Classification of Astrocytomas with Neural Networks. Path. Res. Pract. 192:
354 (1996)

2. IGLESIAS-ROZAS, J.R.; GRIESHABER, F.A.; SOBOL, A.:
Superiority of Neural Networks over discriminant Analysis Methods for
Histological Classification of Human Meningiomas. Electronic Journal of
Pathology and Histolology 2: 961-05 txt (1-11). (1996)

3. IGLESIAS-ROZAS, J.R.; BRODBECK, A.; ZELL, A.: Classification
of the Oligodendrogliomas with neural networks. Clin. Neuropathol. 15: 275
(1996)

4. IGLESIAS-ROZAS, J.R.; GRIESHABER, F.A.; SOBOL, A.:
Superiority of Neural Networks over discriminant Analysis Methods for
Histological Classification of Human Meningiomas. Electronic Journal of
Pathology and Histolology 2: 961-05 txt (1-11). (1996)

5. IGLESIAS-ROZAS, J.R.; BRODBECK, A.; ZELL, A.: Classification
of the Oligodendrogliomas with neural networks. Clin. Neuropathol. 15: 275
(1996)

6. BRODBECK; A.; IGLESIAS-ROZAS, J.R. ; ZELL, A.:
Classification of Astrocytomas with Neural Networks. Electronic Journal of
Pathology 2: 963-05 txt (1-13) (1996)

7. BRODBEC, A.; IGLESIAS-ROZAS, J.R.; ZELL, A..:Classification
of Oligodendrogliomas Using Neural Networks. In: Classification and
Knowledge Organiztion. R. Klar -O.Opitz (Eds.) Springer Verlag. Berlin 1997
pp: 434-440.

8. IGLESIAS-ROZAS, J.R.; BRODBEC, A.; GRIESHABER, F. A.; SOBOL
A.;.; ESPARZA, J.; SANCHEZ B.; GRIFF S.; KUNSMANN, J: Praktische Bedeutung
neuronaler Netze in der Pathologie. Verh. Dtsch. Ges. Path. 81:733 (1997)

9. IGLESIAS-ROZAS, J.R., ESPARZA J., SCHERF M.: Diagnosis of
human Oligodenrogliomas with the Help of neural Network "NeuroShell Easy
Classfier *" . Acta neuropathologica 99/4:445-446 (2000)

10. IGLESIAS-ROZAS J.R., ESPARZA J., SCHERF M.: Diagnosis of
human Oligodenrogliomas with the Help of neural Network "NeuroShell Easy
Classfier *" Analytical and Quantitative Cytology and Histology
2000;22:383-392

Monitoring Cardiovascular Systems
Pacific Northwest National Laboratory used Ward System Group neural networks to develop a training algorithm that receives physiological data such as heart and breathing rate from a monitor worn by an individual and develops a model of the person's cardiovascular system. This model may be used to predict what a person's response would be in a particular situation (such as fighting a fire or becoming involved in police work) or to predict a changes in an individual's health over time.

Monitoring Physiological Signals
The University of Colorado School of Medicine used NeuroShell to develop an application consisting of four neural networks which detects patient breathing abnormalities or malfunctions in equipment when a patient is under anesthesia.

Determining Drug Significance
Ross Donolow of Zeneca Pharmaceuticals in Wilmington, DE, developed a neural network application to examine data from drug experiments which reduced analysis time by 68 percent. The neural network was trained to detect "good" bladder pressure peaks from hundreds of peaks in data files. These peaks are critical for determining drug significance in experiments designed to screen more drug compounds.

Mr. Donolow previously used the paper and pencil method to analyze these peaks from strip chart tracings, a method that took 40 hours per month. Using Ward Systems Group networks reduced analysis time to 13 hours per month.

Forecasting Treatment Costs
The doctors and administrators from the Cleveland Clinic developed a neural network which predicts the cost associated with heart catheterization and intervention. The neural network predicts whether the patient will be a normal or high cost case. The inputs include the patient's age, sex, and other medical problems as well as structural data on a patient's heart and blood vessels.

Forecasting Length of Patient Stay
A team from Johns Hopkins University School of Medicine developed a neural network to forecast which patients in the hospital's Surgical Intensive Care Unit would remain seven or more days. The neural network's inputs include age, the use of drug or other therapies, and physiologic signals such as heart rate and blood pressure, and laboratory results such as white blood cell counts.

A Hopkins study concluded that the "three neural networks were substantially better predictive models than the multiple linear regression model...." In a similar application, a team from St. Michael's Hospital in Toronto, Canada, devised a neural network to forecast a patient's length of stay following cardiac surgery.

Diagnosing Prostate Cancer
Kaman Sciences Corporation (Colorado Springs, CO) used Neuroshell to create a network which can predict prostate cancer. This system was featured on CNN and in the Wall Street Journal as a major breakthrough. NEural networks were trained to predict biopsy results based upon PSA (prostate specific antigen) tests and to predict cancer recurrence after treatment.

Prediction of Cardiac Surgery Mortality and Predicting the Risk of Malignancy for Mammographic Abnormalities
Richard K. Orr, MD, MPH has completed an impressive number of studies over the years in which he successfully used neural networks from Ward Systems Group, Inc. to predict cardiac surgery mortality, and to predict the risk of malignancy for mammographic abnormalities. Dr. Orr is at the Department of Medical Education, Spartanburg Regional Healthcare System in Spartanburg, South Carolina. In his studies he used Probabilistic Neural Networks (PNN), which are found in the NeuroShell Classifier (the genetic method) and in the older classic NeuroShell 2.

Dr. Orr's studies have been presented in a variety of formats:

San Francisco, CA. Jan. 1995. Society of Critical Care Medicine."Use of artificial neural networks to predict ICU length of stay in cardiac surgical patients." [Poster]

Washington, DC. July 19, 1995. World Congress on Neural Networks. "Use of a probabilistic neural network to predict mortality following cardiac surgery."

Tempe, AZ. Oct. 14, 1995. Society for Medical Decision Making. Short Course: Neural Networks: Promise and Pitfalls.

Tempe, AZ. Oct. 14, 1995. Society for Medical Decision Making. "Use of a probabilistic neural network to predict survival after cardiac surgery." [Poster]

Boston, MA. Dec. 3, 1995. 1995 International Conference on Health Policy Research (Methodologic Issues in Health Services and Outcome Research). "Using risk-adjusted outcomes to assess mortality after coronary artery bypass graft surgery."

Dixville Notch, NH. Sept. 29, 1996. New England Surgical Society, Artificial Intelligence in Surgery. "Use of an artificial neural network as an aid in mammographic interpretation."

Toronto, ON. Oct. 16, 1996. Society for Medical Decision Making. "An artificial neural network to aid in mammographic diagnosis can be trained with natural language. [Poster]

Lebanon, NH, November 1996, New England Cancer Society. "Use of an artificial neural network as an aid in mammographic interpretation."

Boston, MA. October, 1998, Society for Medical Decision Making. Cost-effective management of nonpalpable breast abnormalities: Use of a computerized model to determine need and type of breast biopsy. [poster]

"Use of a probabilistic neural network to predict mortality following cardiac surgery." Proceedings of the 1995 World Congress on Neural Networks. 1995; II:754-757.

"Use of a probabilistic neural network to estimate mortality risk after cardiac surgery." Med Decis Making 1997; 17:178-185.

Prediction of Obstructive Sleep Apnea From Clinical Criteria
Dr. Wayne Danter and his colleagues at the University of Western Ontario used the NeuroShell Classifier to develop a predictive model for diagnosis of obstructive sleep apnea (OSA). Their GRNN model was able to accurately rule in OSA from clinical data, and GRNN did not misclassify patients with moderate to severe OSA. In their study, they further found that use of the neural network could have reduced the number of polysomnography (PSG) studies:

"Neural Network Prediction of Obstructive Sleep Apnea From Clinical Criteria." CHEST August 1999; 116:409-415.

A Neural Network/Genetic Algorithm Model to Predict Caesarean Section in a Busy Labor Ward
Dr. Ken Giuffre, Director of Anesthesia Research, and his research team at Hackensack University Medical Center of UMDNJ, have constructed a model that predicts caesarean sections on a busy labor ward. Their paper was presented at the 1998 meeting of the American Society of Anesthesiologists. Click here to view.

Prediction of Patient Preferences
Avi Amin, M.D., and Bob Nease, Ph.D., have used NeuroShell Trader and NeuroShell Classifier to predict patient preferences for medical treatments and outcomes. The neural networks used clinical data about the disease as well as information from patient questionnaires on symptoms. The neural networks were very successful at predicting patient preferences. The neural networks performed substantially better than traditional modeling techniques.

The patient preference data are used in decision models which help determine which treatment option to pursue for a given disease. Understanding patient preferences is important because they are critical inputs in the decision models.

The NeuroShell Trader? Sure, it doesn't have to be used for trading, because it is a very good time series prediction system in its own right. However, we happen to know that Dr. Amin, like many physicians we know, is also a trader, so his software serves two purposes!

Dr. Gordon S. Doig has produced an outstanding and comprehensive study comparing a very popular statistical method and our neural networks
His study completed in April 1999, Severity of Illness Scoring in the Intensive Care Unit: A Comparison of Logistic Regression and Artificial Neural Networks, was Dr. Doig's doctoral thesis at the University of Western Ontario, London, Ontario, Canada. Dr. Doig used genetic adaptive and backprop neural networks to make ICU outcome predictions. Rarely have we seen such a complete work. Dr. Doig is now at the London Health Sciences Centre, and we have reproduced the abstract of his work in the news area of our web site. Click here to view.

NeuroShell Classifier Model Could Lead to More Effective AIDS Drugs
A neural network model developed at the University of Western Ontario with the NeuroShell Classifier has shown great potential in identifying drug compounds that may be effective against the HIV1 (AIDS) virus. Use of this model could potentially screen out ineffective drugs prior to expensive human testing and insure that drugs that undergo human testing are more likely to be effective. This result is significant because although there are currently over 30 million people infected with AIDS worldwide, less than 10% can afford effective treatment.

The neural network model used the chemical structures of anti-viral drugs to predict the anti-HIV1 activity in living organisms. The study was made feasible due to the existence of a large database of molecular structures of compounds that have been tested for anti-HIV1 activity. The study used the PNN-based genetic training method in order to use all of the data. The genetic training method uses a one hold out training and testing strategy so that the test case was never included in the training set.

The results are published in a paper called “A Hybrid Classification Tree and Artificial Neural Network Model for Predicting the In vitro Response of the Human Immunodeficiency Virus (HIV1) to Anti-Viral Drug Therapy”. Co-authors of the paper include W. R. Danter, D. Gregson, K. A. Ferguson, M. R. Danter and J. Bend of the Departments of Medicine and Pharmacology and Toxicology, University of Western Ontario, London Ontario, Canada. The paper was published in the Proceedings of the ANNNIMAB-1 Conference, Artificial Neural Networks in Medicine and Biology, held in Goteborg, Sweden, from May 13-16, 2000.

Since the company began in 1988, we have learned of customers who have created applications ranging from modeling chemical reactions to the development of an "artificial nose"!

Following are brief descriptions of some of these applications.

NeuroShell Predictor for Hydrologic Modeling
Dr. Ward Huffman has successfully used the NeuroShell Predictor for hydrologic models, including the forecasting of flood events. Steve Ward received the following email from Dr. Huffman"

"Over four years ago I talked to you about the feasibility of using a Ward Systems Neural Network for my dissertation. You assured me that the NeuroShell Predictor would do the job. Well you were right. The performance of the NeuroShell Predictor far exceeded my expectation and those of my committee chair and committee members. I believe that one of them is now using a Ward Systems product in his own research.

To summarize: My dissertation is done, my doctorate degree is on the wall, I have written three articles on the research that have all been published. The fourth article will be presented to and published by the Intelligent Systems Design and Applications Conference in Rio de Janeiro, Brazil on October 24, 2007."

High-Resolution Thermogravimetric Analysis in the Compositional Characterization of Woody Biomass
Scientists at the State University of New York have used the NeuroShell Predictor to develop a High-Resolution Thermogravimetric Analysis method to estimate the moisture, hemicellulose, cellulose, lignin and ash content of woody biomass. A novel aspect of this work which significantly improves the sensitivity of this technique is that thermal degradation experiments are conducted in an atmosphere of air rather than nitrogen or other inert gas. This minimizes the occurrence of "char" which complicates component resolution especially for the lignin fraction which degrades at relatively high temperatures. A algorithm based on Neuroshell Predictor 2.0 is used to predict the % hemicellulose, % cellulose and % lignin content of woody materials from HR-TGA data and, in most cases, each compositional parameter can be determined within +/- 2% of its "published" value.

Jacob Goodrich, Arthur J. Stipanovic*, and Patrick Hennessy
State University of New York, College of Environmental Science and Forestry,
Faculty of Chemistry and Cellulose Research Institute, 123 Jahn Laboratory,
Syracuse, New York, 13210

GeneHunter in Functional Genomics, Biochemistry and Nutrition
Christian Ettenhuber and his colleagues at the Technical University of Munich have successfully applied GeneHunter in the deconvolution of nuclear magnetic resonance spectroscopy data, permitting a new way of elucidating intermediary metabolic processes in Drosophila melanogaster and other organisms. We have no idea what that means, but we did find out from Dr. Ettenhuber that Drosophila melanogaster is a fruitfly, made famous a few years ago in California. That much of the study we now understand!

You can access the research article in the Proceedings of the National Academy of Science via:

http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=pubmed&dopt=Abstract&list_uids=15096588

However, unless you are a scientist in the field, you may want to read Dr. Ettenhuber's synopis by clicking here.

Meatballs
Jixian Zhang of the Jiangsu University of Science and Technology in China and G.S. Mittal of the University of Guelph in Ontario, Canada have trained artificial neural networks (ANN) to make predictions of the temperature, moisture, and fat content of meatballs during the deep-fat frying processes. Among some of the numerous input variables were: frying time, meatball radius, initial moisture content, initial meatball temperature, and oil temperature. The neural network, using our NeuroShell software, was trained and verified to produce outputs including: temperature at meatball geometric center, average temperature of meatballs, average fat content of meatball, and average moisture content of meatball. They found the hidden neurons behind the neural network to produce minimal error in generating outputs, and were pleased to ascertain via NeuroShell which variables had the most and least influence within their predictions. As a result of their research, they found that “trained ANN can be used to predict outputs . . . with high precision.” They concluded, “predicting the meatball deep-fat frying process by ANN using simulation data is a simple, convenient and accurate method. ... The ANN model can be used in food process control for its accurate prediction results and real-time fast response.” Their paper, “Prediction of Temperature, Moisture and Fat Contents in Meatballs During Deep-Fat Frying Using Artificial Neural Network” was published in the Proceedings of the 99 International Conference on Agricultural Engineering in Beijing, China, December 1999.

Our question: What about the spaghetti?

Artificial Neural Networks for Automatically Estimating the Age of Fish.
"The ageing of fishes, and consequently the determination of their growth and mortality rates, is an integral component of modern fisheries science." This quote from another researcher begins Simon Robertson's paper that he wrote with his colleague Alexander Morison at the Marine and Freshwater Resources Institute in Queenscliff Victoria, Australia. Robertson and Morison's paper "A trial of artificial neural networks for automatically estimating the age of fish" is the first of several papers from them on fisheries science. This one details their experiments on determining the age of fish by pattern recognition using our neural networks of "sectioned otoliths". (Our layman's translation: they sliced up fish and made bitmap images by illuminating the thin slices under a microscope!)

Robertson and Morison concluded "This study demonstrated that ANNs have the ability to estimate the age of at least some species and some age classes of fish from images of their otoliths with an accuracy comparable to that obtained by an experienced reader. That the ANNs used in this study, when provided with input data in the most simplistic form (a single array of luminance values), were able to provide acceptable age estimates strongly supports the potential of the method."

In their second paper, Robertson and Morison confirmed their earlier work and examined a specific neural network type that is a favorite of many of our users: Probabilistic Neural Networks (PNN). Their work became Chapter 19 in the book Ecological informatics, edited by F. Recknagel, and is titled "Age estimation of fish using a probabilistic neural network." One of Robertson and Morison's results was "The fitted neural network models achieved R squared in excess of 0.8 for all species indicating that they explained a high level of the variation within the datasets." They were also able to identify the most relevant input variables using the "contribution factors" that our software provides. Another conclusion was that the PNN networks showed improvement over the results of backpropagation models.

Diagnosing Turbine Engine Problems in Tanks
Pacific Northwest National Laboratory and the U.S. Army Ordnance Center and School used Ward Systems Group neural networks to develop a neural network that diagnoses faults in the fuel system of an M1A tank. Correct diagnosis could prevent an automatic shutdown of the engine, which could be crucial to battlefield success.

Monitoring Product Quality
The state of Florida used NeuroShell in order to comply with a Food and Drug Administration requirement for the processor to list on the container the country or state of origin for each juice lot.

The neural network's training patterns were taken from a data base that analyzed orange juice samples from different states and countries for the presence of 15 minerals, such as sodium, calcium, potassium, etc. The scientists used inductively coupled plasma atomic emission spectrometry to measure the amounts of minerals in each sample.

CTS Corporation in Brownsville, Tex., uses NeuroShell to monitor the quality of stereo speakers as they come off the assembly line.

Identifying Chemicals with an Artificial Nose
Pacific Northwest National Laboratories used Ward Systems Group neural networks to develop an "artificial nose" that operates by blowing a chemical vapor over a sensor array. The sensor signals are then fed to the neural network, which identifies the chemical. The nose may be used to identify chemicals in hazardous environments or to monitor air quality. Other potential applications include using the nose to evaluate body odors to identify possible medical problems and monitoring quality in the food industry.

Modeling of Chemical Reaction Kinetics
A chemical research center has used Neuroshell to model the rate of chemical reactions when the temperature and the amount of catalyst are altered. The center is examining other suggested applications, including monitoring and controlling a chemical processing plant, predicting properties such as strength and elasticity for a chemical mixture based upon inputs, and forecasting sales.

Zinc Coating Life Prediction
Galvanizing, which produces a zinc coating on steel surface, is one of the most effective methods for corrosion protection of steel. Like other metals and alloys, galvanized zinc coatings corrode at certain rates depending on the environmental conditions. The corrosion rates of zinc coated steels in atmospheric environments have a wide range, about two orders of magnitude. It is important to know the specific corrosion rate in a given application environment in order to effectively use zinc coated steels in outdoor structures.

Traditionally, a common method for estimation of the life of galvanized steels has been the use of a generalized value for the different types of atmospheres known as rural, industrial, urban and marine. However, this non-specific approach is no longer adequate to meet the demands of the market place. Today, the users of galvanized steels are increasingly asking for information on performance certainty, that is, on life prediction. Also, as products are becoming more application-specific, more relevant information on corrosion rates is required, which calls for more accurate prediction methods.

A new prediction method was developed using statistical methods, neural network technology and an extensive database by Dr. X.G. Zhang at Cominco Product Technology Centre. The neural network models were developed using NeuroShell. The models take the basic and readily available weather and air pollution information as inputs and produces corrosion rates as output. The method is now being utilized to develop Internet based prediction software.

Chicken Feed
Yes, literally. Dr. William Roush and his colleague Terri Cravener at the Penn State Department of Poultry Science are pioneers in applying neural networks to analyze poultry feed and other aspects of the health of "broiler" chickens. Their extensive research has placed them in the forefront of the scientific application of non-linear modeling in their field. Dr. Roush also speaks about and teaches neural network applications. Among their many papers are the following:

Roush, Kirby, Cravener, and Wideman. 1996. Artificial Neural Network Prediction of Ascites in Broilers. Poultry Science 1996, 75:1479-1487.

Roush, Cravener, Kirby, and Wideman. Probabalistic Neural Network Prediction of Ascites in Broilers Based on Minimally Invasive Psychological Factors. Poultry Science 1997, 76:1513-1516.

Roush and Cravener. Artificial Neural Network Prediction of Amino Acid Levels in Feed Ingredients. Poultry Science 1997, 76:721-727.

Cravener and Roush. Improving Neural Network Prediction of Amino Acid Levels in Feed Ingredients. Poultry Science 1999, 78:983-991.

Roush and Wideman. Evaluation of Broiler Growth Velocity and Acceleration in Relation to Pulmonary Hypertension Syndrome (PHS). Received for Publication April 5, 1999.

Improving Drinking Water Quality
In order to improve drinking water quality while reducing operating costs, many drinking water utilities are investing in advanced process control and automation technologies. The use of artificial intelligence (AI) technologies, specifically artificial neural networks (ANNs), is increasing in the drinking water treatment industry as they allow for the development of robust non-linear models of complex unit processes.

Chris Baxter and his colleagues at the University of Alberta, Department of Civil and Environmental Engineering, have been applying NeuroShell neural networks to water quality modeling as well as drinking water treatment process modeling. They have done some impressive work in several case studies at two large-scale water treatment plants in Edmonton, Alberta.

Drug Side Effects Prediction
One young scientist (age 15!) used neural network models to predict the side effects of new drugs based upon the known side effects of existing drugs. A Frederick High School freshman won first place in the 2000 Frederick County Maryland Science Fair, Computer Science Division, with her project. She coded the characteristics of 60 known and widely used drugs as inputs to her models. She then applied her models to 20 new drugs. She built a classification model for each side effect (e.g. headaches) using the NeuroShell Classifier. Since she had a small training set, she needed the "one-hold-out" feature of the genetic method in the NeuroShell Classifier to be able to use all of her 60 drugs for both training and evaluation. The student averaged about 88% accuracy in her models.

In addition to winning first prize, the project also received the Army Achievement Award and a $200 special award from Intel.

Stock Market Modeling by Individuals
Jeff Parent has had success in modeling the US stock markets using the NeuroShell Classifier. He collects technical analysis-based data taken at the outset of an outlier event and uses the data as inputs for a classifier neural network. Many of the trained neural networks correctly identify close to 70% of the outcomes from the validation set. This is a good result considering the notoriously noisy stock markets. Parent has been analyzing stocks for over 12 years and has authored articles for Technical Analysis of Stocks and Commodities Magazine.

Model Creation

Many analysts and academics do not believe it is possible to make profitable, short-term predictions of the stock markets. It is often said that markets are efficient in adjusting price to reflect changes in value thereby providing few trading opportunities. In general, Parent agrees with this statement. However, by focusing on outlier events only, he has been able to increase the odds of making successful predictions. At the beginning of an unusual price movement, more so than in a normal market, he believes there are clues to a potential outcome embedded in the technical factors of a stock. Working with a classifier neural network is natural because the results he is looking for are in the ‘event will/will not likely occur’ form.

An example of an outlier event is a stock trading at an unusually low level from a previous day’s average price. The outcome to be determined is whether or not the stock will revert back to the average price on the following day. As one possibility, Parent defines a low level as a stock dropping in value by a percentage of its average volatility. Technical indicator values like RSI and relative volatility and general market conditions are collected on the day before the drop. Knowing the condition of the markets puts the stock’s movements into perspective. Equally important is knowing how the stock has acted over the past several days, so recent values of the same technical indicators are included. In all, about 10-12 values are known at the moment before the event. The final piece of data collected is the outcome. After the ‘dip’, did the stock revert back to the average price or not?

Parent describes the process, “Up to ten years of data from actively traded stocks are scanned to net about 40,000 ‘dip’ events. Once collected, the event data must be divided into training and validation sets. Events are sorted by date and separated into somewhat randomly selected chunks of 6-12 month segments. One portion is set aside for validation. The balance is used to for training purposes. Having equal amount of both outcomes in the training set is necessary to ensure the neural network finds a non-trivial solution. Therefore, equal amount of data from positive outcomes (the stock reverts to the average price) and negative outcome are added together to complete the training set.

Validation and Usage

He continues, “I use the Classifier’s TurboProp2 method. It trains faster and I intuitively believe it is more likely to be aligned with the task. To be considered a success, I look for low false-positive rates and consistency between training and validation results. Normally, the ‘dip’ buyer strategy requires setting price alerts or placing limit buy orders on up to 500 stocks each day. Due to capital limitations, positions are taken only in the first 10-20 stocks hitting their targets. Using the trained neural network the same list is screened. A smaller watch list is created and profit on each position taken is potentially better. I have been very happy with the results. The Classifier is easy to use and very powerful.

Sales Revenue Prediction
The Resource Planning Consultants of KPMG Peat Marwick created a neural network model using Neuroshell which predicts the average monthly sales revenue for a client’s automobile maintenance center franchise. The model was successful in spite of the fact that the franchise was a new one, which meant few sites were available to provide data.

Predicting Risk of Bankruptcy
Oklahoma State University created a neural network which predicts the risk of bankruptcy for various companies based upon ratios such as working capital/total assets, retained earnings/total assets, earnings before interest and taxes/total assets, market value of equity/total debt, and sales/total assets. The neural network model was more successful than a similar model which used discriminant analysis.

Market Forecasting and Issue Selection Decisions
Since we started selling the first version of NeuroShell in 1988, thousands of our users have built stock and commodity market models. Most are small investors, and you can read interviews we did of several of them.

However, one Fortune 100 company, which makes vehicles familiar to everyone, uses our neural networks to make S&P 500 futures trades worth many millions of dollars. The neural networks worked so well on the S&P that the company has opened other funds using our neural networks. This company uses eight simple ratios as inputs to the neural networks to get about 70% of their trades correct.

Smaller companies have reaped the benefits of our technology as well. One small capital management company was on Inc Magazine's list of the 500 fastest growing companies three years in a row, thanks to their neural network predictions. The technical division of the company split off and was quickly acquired by State Street Global Advisors, who made the CEO the Research Director of State Street. The company still uses our neural networks for thousands of financial predictions.

Academic Thesis Comparing NeuroShell 2 to Regression for Market Prediction
As a student at the US Naval Post Graduate School, Lt. Jason Kutsurelis produced a thesis showing not only that neural networks can predict the market, but that they also outperform regression at the task. Lt. Kutsurelis' thesis is a landmark piece of work that is sometimes copied from our website and cited by other producers of neural network software, even though he used NeuroShell 2 to do the work.

Although this is an extremely technical paper, you can read it by clicking here. If you don't want to wade through the math and technical details, the abstract at the beginning is an easy to read summary. We appreciate Lt. Kutsurelis making this work available to us.

Many customers have used Neuroshell and GeneHunter for making sales forecasts, predicting customer arrival numbers, selecting target markets, and process control applications.

Forecasting Sales of High Volume Consumer Goods
As a consultant to Nestlé Italy, Filippo Vasta from Mbym s.r.l. developed a GeneHunter application that is an adaptive exponential smoothing algorithm for forecasting sales of high volume consumer goods such as chocolate bars and cookies. His model exceeded 96 percent accuracy for the year, with average monthly accuracy for the twelve coming months ranging between 80 and 95 percent.

Vasta used GeneHunter in an Excel spreadsheet to develop the application. According to Vasta, "the innovation comes from searching for seasonal adjustments, the initial value of the time series, and the promotional plan." His model separates the time series into a baseline component derived from trends and seasonality and an incremental sales component.

Vasta says that the user does not need to make any assumptions about promotional efficiency as this is calculated by the model as a result of and in the context of the calculation of all other parameters.

"Once all the data is calculated, the forecast is made by projecting the baseline and calculating the future incremental sales based on assumptions for the promoplan for the coming periods, i.e., the proportion of customers interested in the promo each month for the coming year," according to Vasta. He added that the forecast is clearly affected by the accuracy of the data.

Vasta said the same algorithm can be used for optimizing promo spending plans over a range of SKU's (Stock Keeping Units) within a product family.

For more information about this forecasting model, please contact Filippo Vasta, MbyM s.r.l., at info@mbym.it.

Optimizing Pricing Strategy
Dr. James Dauer, professor of computer science and applied mathematics at Elmhurst College in Illinois, has developed a unique and highly effective methodology for product pricing using GeneHunter. The model incorporates pricing guidelines for each product that allow the user to set intra-product line price ratios, max/min limits, and volume discounts. In addition, built-in price elasticity curves dynamically adjust product volumes as a function of the changing prices to reflect expected demand at each price level. The model then selects that combination of prices across all products yielding the highest profit -- an intractable task using any traditional optimization technique. In addition, the model estimates the unit volume for each item based on the resultant optimal pricing strategy. Application areas include: all “big box” retailers, restaurants, catalog sales, or any situation involving a myriad of pricing combinations across hundreds, or even thousands, of products. It is ideal for estimating the impact of price changes on product demand. Dr. Dauer’s paper entitled, “Survival of the (Pricing) Fittest” recently published in the Journal of Pricing Professionals, illustrated how a fictitious fast-food chain was able to increase profit by $4.5 million across 100 stores using his model. The actual model formulation is proprietary. Dr. Dauer may be contacted at jimd@elmhurst.edu, or (630) 617-3124, for further information.

IntelliSearch Uses NeuroShell Classifier to Fuel Yamaha Motorcycle Sales in Brazil
The Brazilian company IntelliSearch recently inked a deal with the Yamaha Motors branch in Brazil (Yamaha Motor do Brasil or YMDB for short) to provide credit analysis for new dealers. Carlos Nogueira built the scoring application with NeuroShell Classifier to replace an existing rule based tool. The project also includes the development of an application that will consolidate and validate financial statements sent by YMDB dealers on a regular basis. The application will select and pre-process some of the data from these statements, creating many inputs that will feed the credit analysis module.

Nogueira had been working as an IT consultant with Yamaha on other projects and made a contact with the person responsible for credit analysis. "He showed me the current credit scoring tool, in fact a credit risk classification tool, which was an internally developed Excel spreadsheet with a few worksheets and macros. As it usually happens with this kind of tool, it became outdated in terms of inputs, rules, and weights." Nogueira said the system was developed by a YMDB employee who works in another division and can no longer work on the model. Moreover, the old scorecard model wasn't able to cope with the quick market changes faced by Yamaha, such as the lack of accurate financial statements data from many dealers requiring a shift to new input data sources.

Nogueira won the contract for IntelliSearch by building a prototype for YMDB with the Classifier. Nogueira jumped in and made a PowerPoint presentation that showed the advantages of using neural networks for credit classification. "I got their attention and the person responsible for credit analysis got hooked on the neural network appeal. I bought your NeuroShell Classifier and built a prototype in Microsoft Access. I gave a demonstration and it was a wild success. As a plus, the prototype also performed the same conventional 'scorecard like' procedure done by YMDB’s Excel tool, so I was able to compare the two approaches and show them the neural network pros. At the end of the demo, they were very excited and told me that it was exactly what they were looking for."

Nogueira’s company had previously developed credit scoring applications for four other clients using other neural network tools, but Nogueira switched to the NeuroShell Classifier for the YMDB project because "of a few important reasons. The first one is that the NeuroShell Classifier core engine and user interface are more up to date than our former choices. Second, the NeuroShell Classifier tool adds genetic algorithms for input optimization during training. Third (as an outcome of the second), the Classifier is able to show the relative importance of inputs, a must have to YMDB and a key differentiator in the NN based applications arena."

YMDB has experienced rapid growth for the past three years, with a 30% annual growth rate for gross sales. Some of that growth comes from improved technology in the Yamaha motorcycles, but most of the growth emerged from an aggressive new dealer certification program. Three years ago there were 200 dealers and today that number grew to nearly 350. More dealers, however, meant more dealers who were delinquent in their payments. YMDB hopes to alleviate this problem by adding an improved credit analysis modeled with the NeuroShell Classifier as part of the dealer certification process.

IntelliSearch (http://www.intellisearch.com.br/) is a small local company but with a lot of "IT grown gray hair," according to Carlos Nogueira. In addition to credit scoring applications, the company offers data security and cryptography related services in partnership with other companies. IntelliSearch is also starting a Project Risk Management activity.

Predicting Consumer Response
Predicting how many consumers respond to promotional offers such as the ones that appear on cereal boxes is such a high volume business that companies no longer do it for themselves. They hire promotion fulfillment companies that do the job for many of the nation's largest consumer products companies. In addition to shipping the items such as T-shirts, they offer their client's guidance on how to structure their promotion.

The promotion company turned to GoalAssist Corporation in Minneapolis, a consulting group experienced in the use of pattern recognition tools, for assistance in improving the accuracy of consumer response predictions. Jerry Hammann of GoalAssist thought that a neural network was the best tool for solving the problem because the companies had historical data on previous customer responses.

Hammann created a prediction system that used two neural networks. He trained the first neural network with the NeuroShell Classifier. Inputs such as media type, circulation, product type, number of UPCs required, refund or promotion amount, etc., resulted in four broad ranges of response to the promotion.

The second neural network used the original inputs from the first neural network plus the outcome classification from the first neural network. It was built with the NeuroShell Predictor to forecast the response rate. The average error of the prediction neural network was 3.5%. "The promotion fulfillment company views these models as a competitive edge," according to Hammann.

Using GeneHunter in Adaptive Controls
One of our long time users, Dr. Jan R. Schnittger of Nice, France, has kindly allowed us to publish on our site his paper which describes how GeneHunter can be used to model adaptive control processes. Dr Schnittger is a retired professor of the Department of Machine Design at the Royal Institute of Technology, Stockholm, Sweden. According to Dr. Schnittger, "Such a model may be part of an adaptive control or it could detect process failures. In other cases it could replace the actions of a skilled operator." Click here to read Dr. Schnittger's paper entitled "A Genetic Algorithm in Adaptive Controls". Dr. Schnittger can be contacted at schnittger@writeme.com.

Student Thesis Uses Neuroshell to Profile Web Sites
Ryan MacDonald and his thesis supervisor, Dr. Daniel L. Silver of Acadia University in Nova Scotia, have finished a groundbreaking project in which they have taught neural networks to profile the Internet. We were quite impressed at the level of this work, which was only Ryan's bachelor's thesis! Ryan (mailto:035316m@acadiau.ca) has written a condensed paper from his thesis which can be read by clicking here.

Student Wins Science Fair with Sales Prediction Project
A high school student swept the honors in the Computer Science category of the Frederick County, Maryland, 2001 Science Fair. The judges awarded the project "Paging Neural Networks, Company Statistician" with first place in the category, a U.S. Army recognition award, a U.S. Air Force Award, and an Intel Excellence in Computer Science Award of $200.

According to the student "I presented the project to about seven judges. Many of the judges have worked with neural networks before, and they seemed to like my project very much."

The project used both the NeuroShell Predictor and regression analysis to do sales predictions for the student's web-site business. The business, http://www.virtue.nu/moonslush, caters to teenage girls in need of affordable lip gloss and other teenage accessories.

The following quotes were taken from the student's project:

"How accurate can a neural network predict sales for my Internet start-up Moon*Slush? What independent variables most influence the amount of business generated, according to the neural network? How do the predictions for the non-linear neural network compare to the predictions of a regression analysis?"

"I was able to get an optimal correlation of 0.844761, and, in general, variables pertaining to the events and conditions of the days are the best variables to use to create the best neural network for sales prediction. These type of variables can include the season of the year, weather (temperature, precipitation, pollen count, etc.), the number of holidays, distinctions between days of lesser stress, consideration of outside, monumental events to the target demography (concerts, sport events, Academy awards, season finales, etc.), events at other websites, and so on."

The student also experimented with other variables such as colors in the website, advertising, and number of items for sale.

"Compared to linear regression, neural networks have a stronger accuracy if the data is non-linear; simple linear regression models do not take multiple variables into account. A neural network also does not have to predict in a linear path; the linear regression models are restricted to only predicting in a linear route, which therefore increases the amount of error if the data is non-linear."

"From my experience with neural networks, other small companies may find neural networks as a great solution for sales prediction and company optimization. This method is an affordable and easy way to predict sales without the hassle of having to hire a full-time statistician. Neural networks also forces an entrepreneur to discover information about his business; the knowledge that is gained from the use of neural networks to predict sales is priceless, as it can dynamically improve one's company for years to come."

Selecting Audit Targets
The University of Illinois is working on a neural network application which identifies accounts to audit based upon university accounting records. VISA Corporation is leading the way in this area, having developed an extensive fraud detection system which employs a neural network.

Predicting Fish Catch
The National Marine Fisheries Service used NeuroShell to develop a model that predicts the landings for Atlantic menhaden fish with less than a 2 percent average error. The previous model, which used multiple regression analysis, averaged an 11 percent difference between model forecast and actual landings.

The model’s predictions of fish landings are critical to the planning efforts of the U.S. government, banks, and the fishing industry, which use the numbers for deciding upon the number of fishing licenses to issue, evaluating a bank loan to construct a new vessel, or planning the capacity of a processing plant.

Scheduling Staff Members
The Medical Center of Delaware used GeneHunter to develop a schedule for medical residents. The residents have to work in outpatient clinics, in addition to rotating through different parts of the hospital, such as the coronary care unit. Creating the schedule was further complicated by the fact that students were assigned to different locations depending upon their training level. The GeneHunter schedule had only 6 conflicts for the entire year, while the Center’s "pencil and paper" effort had 24 conflicts!

Forecasting Reservations and Covers
Major corporations in the hospitality industry (theme parks) have created models with Neuroshell to predict the number of reservations or restaurant covers (patrons) in order to effectively schedule enough staff to handle the workload.

Selecting Target Markets
One analyst has successfully used Neuroshell to select candidates to receive mailings concerning environmental and political fund raising activities.

Oil Industry
Most of the major oil companies have used Neuroshell for making financial predictions. Some have used NeuroShell to make predictions regarding manufacturing processes, product composition, and oil exploration sites.

Optimizing Manufacturing Processes
Allen-Bradley Company of Milwaukee, WI, a manufacturer of circuit boards and other high tech products, needed to minimize the time it would take to manufacture different types of circuit boards on an assembly line. They tested a model developed with GeneHunter against a $150,000 scheduling software package targeted at manufacturing operations and found that GeneHunter gave better results. The model they developed in GeneHunter was based upon an example program that is included with GeneHunter.

Predicting Service Calls, Customer Transactions
Brooklyn Union Gas Company used NeuroShell to predict one to two days in advance the number of crew members needed for service calls based upon the time of year, predicted temperature, and day of the week.

Banks have used NeuroShell technology to predict the number of customer transactions likely to occur at different times during the month in order to properly schedule staff. Hotels, restaurants, and switchboards predict the number of arrivals or calls in order to efficiently schedule personnel.

Employee Selection
Neill Carson and Associates of Houston, Tx. uses a variety of neural network configurations in NeuroShell in its employee selection products. Carson’s "Pro" Series, is used by Southwest Bank of Texas in the initial hiring and placement of candidates. “The beauty of neural networks from a psychological perspective is that they allow local norms and patterns to be used in prediction rather than relying on studies done elsewhere that may or may not apply to the local circumstance, “ according to Carson. Carson's data strongly suggest what most managers intuitively know: customer service characteristics may be different in Buffalo than in Dallas. Neural networks allow for "local" validation and prediction studies that would be cost prohibitive (as well as less effective) using traditional methods.

“Neural networks are especially attractive for use with human behavioral and psychological data because they are insensitive to the often non-linear nature of such data, “according to Carson. Another Carson product, "Sales Path", uses neural networks to identify patterns of success in sales employees. Use of the neural network to classify sales people led to the discovery that successful sales people fall into two categories that are mirror images of one another: One pattern of high performance was characterized by high scores in the "traditional" sales traits of sociability, empathy, conformity, and warmth. The second pattern was not especially sociable, low empathy, strong individualism, and aloofness with people. These were the technically oriented sales people, who probably achieved their high performance by being a strong technical resource to the customers. Carson noted that companies could hire underachievers by selecting an “average” of these two patterns.

Classifying Applicants
Wendell Williams of Emergenetics Consulting Group, LLC, in Acworth, Ga., reports a success rate of 82 to 100 percent accuracy in determining probability of whether or not an applicant has the social qualifications to perform the job. (The neural networks do not determine the best candidate for the job, only the applicant’s expected performance level in the job.) Williams has taken 25 years of personality and performance research to determine his inputs to the neural networks.

He begins by giving the ESP test to current employees of the company in the same job family (i.e., engineers, managers, sales, clerical, professional, technical, etc., would all constitute different families). He also asks their supervisors to rate each subordinate in 23 different performance areas using a scale of "not enough", "just right" or "too much".
Based on this information, he creates 23 runtime net files and uploads them to the client’s site.When the client tests a new applicant, he enters scores onto the web site and the program reports back the applicant's expected performance level in each of the 23 areas.

“Using the NeuroShell Classifier has cut my analysis time in half. The Classifier combines three features that I always thought were mutually exclusive: 1) ease of use, 2) a robust application, 3) and low price. Don’t underestimate the power because it looks simple,” says Williams.

NeuroShell Insures Fairness in Evaluating Genetics Nurses
The Center for Self-Sustaining Leadership (CSL) has successfully used NeuroShell to bring consistency in an otherwise subject evaluation process to certify nurses in a genetics specialty.

Credentialing genetics nurses goes beyond academic degrees. The different educational levels don’t address the clinical skills required in specialized fields such as genetic nurses. So in addition to meeting specific degree requirements (a bachelor’s degree for a generalist and a master’s degree or equivalent for a specialist), applicants for the genetics certification needed to submit a portfolio of four cases studies based on their work.

The portfolios are evaluated according to performance standards for the genetics specialty developed by the International Society of Nurses of Genetics (ISONG). CSL added expertise in quantifiable measurements of portfolio components and application of neural networks to ensure evaluator reliability. CSL was required to translate ISONG standards into observable and measurable components of standards and performance indicators. The CSL team found that “working through this process is difficult and most groups struggle with the challenge of deriving performance indicators from standards of practice.”

The portfolios are scored by a panel of evaluators who had been thoroughly trained in the process. It takes an average of one to two hours for an evaluator to score a portfolio. After scoring a portfolio, the evaluator enters his/her scores on a secure web site designed by CSL. The evaluators do not discuss the scores among themselves until all of the data is entered. Once the scores are received, a CSL staff member enters them into a pretrained neural network based upon previous evaluations. The neural network determines consistency of evaluator scoring and makes an initial determination if an applicant meets the standards. If an outlier appears, the evaluators are asked to rescore that particular section. If the problem reoccurs, the evaluators make the final decision. “It is essential to remember that ultimately the evaluators, and not a computer program, make the final decision,” according to CSL.

The neural network has shown an approximately 98% accuracy rate. The neural network insures that evaluator scores are compatible with scores from previous years and current ones. The neural network can still be applied even if scored items are added or removed from the portfolio.

For more information, refer to the book Genetics Nursing Portfolios, A New Model for Credentialing edited by Rita Black Monsen. Chapter 7 entitled “Use of Neural Net Technology to Quantify Portfolio Evaluations,” by Dave Holmes, Robert McAlpine, and Jack Russell, discusses the specifics of the neural network. The book is published by the American Nurses Association.

Universities and private training institutes use Neuroshell and GeneHunter to teach neural network and genetic algorithm application development because of the easy to understand user interfaces built into each product. Students are able to concentrate on creating working applications.

Other universities are using NeuroShell to predict student performance and to monitor student errors in order to improve teaching methods.

Pro Football Betting
Ward Systems Group does not encourage anyone to gamble. Having said that, we know that many of our customers are using our software for football betting (pro in the US and soccer overseas). This is the story of one such user. His name is Paul Smith, and he has been betting on pro football for about 20 years now. What we hope impresses you about Paul Smith is that the only winning season he has had in that 20 years is the current one, during which he has used the NeuroShell Predictor to get his picks.

Paul is retired from the aviation industry, and although he likes numbers, he stresses that he is no mathematical genius and even flunked algebra in high school. He has never had a statistics course and took only business math in college.

Paul is not a total stranger to neural networks, having used NeuroShell 2 to bet on trotters and pacers a few years ago. Although he made a little money at that, he lost interest because the payoff is so low on trotters and pacers.

So when he called last year to upgrade his NeuroShell 2, we suggested he consider buying the NeuroShell Predictor as well. He did, and now he thanks us regularly for the suggestion. Halfway through the season Paul gave us this report, which we placed in the User Comments section of our website:

"I use your software to bet on pro football games. I get 74% winners on a base of 139 games using the NeuroShell Predictor. I just use 5 factors and play 5-7 games a week. I think the software is fantastic."

The other day we heard from Paul again. He is now hitting 80% on the "under and over" games, making 4 to 5 plays per week. He is winning about two-thirds of his "parlay tickets." One week he got 9 out of 10 correct on the parlay ticket (he only lost a "push"). We aren't sure what all that means, but we know that Paul is very happy about it!

Paul's inputs aren't rocket science, but they are well thought out. He takes care to make sure that his numbers make sense. For example, he codes the variables so that they aren't unreasonable at the beginning of the season when few games have been played, and he doesn't bet until he has enough data. He also eliminates games which seem unreasonable in terms of winners or scoring. We told Paul that the terms "normalizing" and "eliminating outliers" are close statistical terms for what he is doing.

We were so impressed several months ago with Paul's method of coding variables that we changed our tip called "Sports Prediction Using the NeuroShell Predictor" and "Sports Prediction Using the NeuroShell Classifier" on our advanced user’s technical support website to reflect what he does. Yes, we can always learn a thing or two from our users!

Former Race Horse Owner Beats the Odds with NeuroShell
Steve Wright of Nacogdoches, Texas, has had a handicapping passion since the age of 13. He was so fascinated with racing that later in life he owned several horses which he ran in claiming and allowance races at the Fairgrounds in New Orleans and Louisiana Downs in Shreveport. He learned a great deal about racing from those experiences, but finally had to sell his horses because they were taking too much time from his successful communications business. When he sold his business in 1995, he was able to get back to his lifelong hobby of handicapping.

Steve knew computers would help him, so he started buying software. He figures he went through about 12 different products in all price ranges as software started becoming available with the emergence of online wagering. Unfortunately, even the most expensive software didn't help him much, so he went about building his own indicator system in Excel, sorting and analyzing past performance data.

Steve concentrated on finding better indicators than the well-known Beyer number. He eventually developed a set of other formulas that outperformed Beyer as he added more information like the amount bet, aging factors, whether the horse disappointed or not, speed ratings, track variants, and other factors.

Then Steve heard about the NeuroShell Predictor. He purchased it and used it at first just to validate that his own formulas were better predictors than others available. In so doing he noticed the "contribution factors" or weightings that NeuroShell computes for each of the variables used in a prediction. He realized that these numbers would be perfect for weighting those variables in his own models!

Now in its ninth revision, Steve's Excel model is doing very well, working best with allowance races, high priced (50K and up) claiming races, and stake races, although it is not very predictable for turf races. Starting in January he took his system to Gulfstream, Aqueduct, and Santa Anita. By the middle of March he had turned an amazing profit of 1470%.