SPE 30202

Predicting Production Using a Neural Network Society of Petroleum Engineers

(Artificial Intelligence Beats Human Intelligence)

Robert J. Boomer, Texaco Exploration and Production Inc.

SPEMember

Copyright 1995, Sooiety of Petroleum Engineers, Ino.

This paper was prepared for presentation at the Petroleum Computer Conferenoe held In Houston, TX, U,S.A., 11-14 June 1995.

This paperwes selected for presentation by an SPE Program Committee following review of Information contained In an abstraot submitted by the author(s). Contents of the paper, as presented,
have not been reviewed by the Sooietyof Petroleum Engineers and are sUbjeoted to correction by the author(S). The material, as presented, does not neoessarily refleot any position of the
Sooiety of Petroleum Engineers, Its officers, or members. Papers presented at SPE meetings are sUbject to publication review by Editorial Committees of the Sooiety of Petroleum Engineers.
Permission to oopyls restrioted to an ebstraot of not more than 300 words. Illustrations may not be oopled. The abstraot should oontaln oonspiouous aoknowledgment of where and by whom
the paper Is presented. Write librarian, SPE, P,O. Box 833836, Rlohardson, TX 75083·3836, U,S,A. (Facsimile 214-952-9435).

Abstract whether or not to drill a well is based, in part, upon

the economic evaluation of the anticipated
Petroleum professionals, that is engineers and performance of the well. A key component of this
geoscientist, routinely make recommendations to drill economic analysis is the predicted production profile,
wells. This process requires the generation of a oil producing rate versus time. This forecasting of the
production profile, oil producing rate versus time, production performance of a new well is one of the
which is a key component in the final economic most important functions a petroleum professional
decision of whether or not to drill the well. The performs. Both underestimating, predicting a
success of the drilling program, and possibly the producing rate that is lower than actual, and
professional's career, rests, to a high degree, upon the overestimating, predicting an oil producing rate that is
accuracy of this predicted production profile. The greater than actual, the producing performance of the
more accurate this prediction the more certain the well is common. Both have a major impact upon the
economic value of the recommended drilling location. actual economic value achieved by drilling, or not
This paper looks at this predictive process as used in drilling, the recommended well. At the extremes,
the Vacuum Field of New Mexico. It explores the either overestimating or underestimating oil or gas
predictive methods used in the past and reports on producing rates results in a potentially disastrous
their accuracy. A new method, using artificial economic decision. Overestimating the production
intelligence, is presented which improves upon the profile will result in the drilling of a well that is not
accuracy ofthis process. economical while underestimating the production
El éxito del programa de perforación, y posiblemente la carrera profesional,
descansa, en alto grado, en la precisión de este perfil de producción previsto.
performance causes a well to not be drilled which
Introduction would have been economical. Usually the result of an
inaccurate production forecast is the drilling of a well
Acting upon a recommendation to drill a well, which whose economic performance does not yield the
usually involves hundreds of thousands if not millions optimum economics for the available drilling
of dollars, is a major decision. The final decision of investment.
Por lo general, el resultado de un pronóstico de producción inexacto es la perforación de un pozo cuyo
desempeño económico no produce la economía óptima para la inversión de perforación disponible.

References at the end of the paper.

195
2 PREDICTING PRODUCTION USING A NEURAL NETWORK SPE 30202
(ARTIFICIAL INTELLIGENCE BEATS HUMAN INTELLIGENCE)

But this is all part of the risk and uncertainty we face

in the oil industry. The industry drilled wells in the VACUUM FIELD
past and will continue to drill wells in the future. We
accept this uncertainty. However, it raises the
question of just how accurate have the forecasts from
the past been and can this accuracy be improved. For
with improved forecasts come greater economic
certainty and greater profitability.

Infill Development Drilling Study Area

A category of significant interest in the Permian Basin

of West Texas and Southeast New Mexico is infill
development drilling. Infill development drilling is
defined as drilling within an existing secondary Figure 1 Field Location
recovery project to reduce well spacing and
consequently the waterflood injection pattern. The field lies at the north end of the Delaware Basin
and Central Basin Platform along a productive east-
San Andres reservoirs in the Permian Basin have west-trending shelf area of the Permian Basin.
produced over 7.7 billion barrels of oil, the majority Production is primarily from the Permian Guadalupian
from secondary recovery projects. This is 46 percent age San Andres formation at a depth of 4400 feet.
of the total production from the entire basin. This The San Andres is composed of cyclical evaporites
large cumulative combined with low recovery, 30 and carbonates recording the many transgressings and
percent of the oil-in-place, makes this reservoir an regressings of sea level occurring around 260 million
excellent candidate for improved development. 1 Since years ago in a climate very similar to the present day
infill development drilling in existing waterflood Persian Gulf 2
projects provides many of the drilling prospects for the
Permian Basin, it was selected as an area to study the Nine secondary recovery projects constitute the
overall accuracy of production performance majority of the acreage, wells, and production from
predictions. Specifically, the Vacuum Field of New the Vacuum Field. Figure 2 shows the outlines of
Mexico is a good area to study since it is a typical San these secondary recovery projects. These projects
Andres field and the individual production cover approximately 32 square miles. The field was
performance of every well in the field is readily discovered in 1929 but field development did not start
available. until a pipeline to the field was laid in 1937.

The Vacuum Field, see Figure 1, is located in the

southeast comer of New Mexico in Lea County at a
distance of approximately 120 miles northwest of the
city ofMidland, Texas.

196
SPE 30202 ROBERT 1. BOOMER 3

Past Methods of Making Predictions

VACUUM FIELD
The Professionals have at their disposal a myriad of
tools to predict the production profile of an infill
development well. Either a deterministic and/or a
statistical method is typically used. cant muy grande

The typical deterministic method for predicting a

production profile is a straightforward process. The
Professional reviews any geological maps, cross-
sections, and existing well logs of the prospect area to
determine overall reservoir quality. Next, the
Professional reviews the offset well's current and
past production performance. Past production
performance includes not just the rates and volumes of
Figure 2 Secondary Recovery Projects
oil production but also the rates and volumes of water
production and water injection. They may also review
Figure 3 is a plot of monthly oil, water, water injection the production profiles of wells from past drilling
and well count from 1970 to present. Also shown is programs. It is characteristic of infill drilling for the
oil cut (% Oil), oil volume as a percentage of total operator to drill wells in programs with each program
produced fluids (oil plus water), and the injection- including from 2 to 20 wells. In the end the
withdrawal ratio (I/W), water injection rate divided by Professionals use this knowledge to generate a
the total fluid volume. The plot shows that over 300 production profile for a specific drilling location.
infilldevelopment wells have been drilled in the field Sometimes a numerical reservoir simulator is used to
since 1970. Field development covers the start of predict a production profile.
water injection in the 1970's and the down spacing of
injection patterns from 80-acre 5-spots patterns to 40- Statistical methods consider the actual variability of
acre 5-spots and to even some 20-acre 5-spots. past drilling programs to generate a single production
profile for all proposed drilling locations, knowing that
averaging the production performance from every
VACUUM FIELD individual well in the drilling program will come close
to this statistically generated profile. The actual
production performance from each well in the most
recent drilling program is first normalized. That is the
oil producing rates are normalized with respect to
initial completion date not calendar date, production
for month number one is the production for the first
month after initial completion and not for the month of
oL:=~~~kJ January. Next these normalized rates are averaged on
1970 1976 1980 1985 1990 this monthly basis for every well in that particular
-G-'" Wells- BOPO/10 BWPO/1 00
drilling program. This resulting production profile is
*aWIPO/100 .... 011 Cut %*10 1jW*1000
used for the next drilling program. Alternatively,
instead of the most recent drilling program the most
Figure 3 Field Performance recent program of "like" wells is used. "Like" in the
fashion of 40-acre wells versus 20-acre wells or

197
4 PREDICTING PRODUCTION USING A NEURAL NETWORK SPE 30202
(ARTIFICIAL INTELLIGENCE BEATS HUMAN INTELLIGENCE)
geologically similar locations, top of structure versus day (BOPD), at the one-, three-, SiX·, and twelve-
down dip. month time intervals.

All of these approaches are successful. Successful in Initial, 3, 6 & 12 Month

that we have used them for decades to recommend Average Cumulative Oil (SOPD)
wells to be drilled and then to have them drilled. Still,
.. PREDICTED
the actual production profile does not match the ,......--------------_=oo
predicted production profile. Many factors that are
1000
. .. . "!........
.
not associated with the drilling program itself will
influence the long-term forecast, such as changing
100 ••
II II •
" "'-"
• ..
..~
-~

injection rates or altering injection pattern. If we 10

constrain the time frame over which we compare the
1~__t__t_l_t_t_+H+-_+__+_i_++++++_____1>__+_+_+_++++l
actual versus predicted production profile, we might
1 10 100 1000
get better accuracy. Monthly production for the first
ACTUAL
year is a good time frame to use since the profitability
• Professional's -Perfect Predictor
of a new well is heavily weighted by the first year's
production profile. Accurate prediction of the rate of
production for just the first year would improve the Figure 4 Professional's Performance
certainty of achieving the predicted economical
performance of any drilling program. Just how Figure 4 shows how accurate the forecasts are by
accurate have these approaches been at forecasting oil plotting the actual performance versus the predicted
production over the first twelve months of a well's performance. This type of plot is called a Confusion
life? Matrix. 3 Included on this figure is a straight line
labeled Perfect Predictor. If you had a perfect
predictor, then all the points would fall on this line
Prediction Accuracy where the actual and preciicted values are equal.
These Professional predictions have a mean error of
esclarecedora
A review of the first year's production predictions 39 BOPD and a mean absolute error of 99 BOPD.
versus the well's actual performance is enlightening. The mean absolute error is an excellent measure of the
In the Vacuum Field of New Mexico over 250 accuracy of the Professional's forecast. For any well
producing wells have been drilled since 1970. The their forecast will average within 99 BOPD of the
Professionals' forecasts were found by searching the actual rate.
well files and correspondence files where Texaco is
the operator or where Texaco has a non-operating Further statistical analysis reveals that the
working interest for the original Authority for Professionals predict within ±20BOPD just 8 percent
Expenditure (AFE) to drill the well. The AFE of the time and within ±1OBOPD only 5 percent of the
typically includes the predicted production profile for time. The Professionals overestimated production
the proposed location. A sampling of 25 of the 250 75 percent of the time. It is obvious that the
wells was used to determine the overall accuracy of Professionals are not perfect predictors.
the predictions. The production profiles were not
biased by a small group of Professionals making One explanation for this poor showing is because of
forecasts nor were they biased by any particular year's the number of variables considered in making the
drilling results. Error calculations were based upon predictions, variables such as oil rate,cumulative oil,
the difference between actual and predicted injection rate, and cumulative injection. The
cumulative production, expressed as barrels of oil per Professionals limit the number of offset wells and the

198
SPE 30202 ROBERT J. BOOMER 5

type of data considered from each well in their A Neural Network initially assumes a random
prediction process. It is very difficult for humans to relationship between all the inputs and the desired
consider more than a handful of variables at anyone outputs. By comparing its first attempt at an answer
time. If a methodology could consider more offset to the desired output, it self-modifies this initial
wells and parameters, such as water production, random relationship into a relationship that best fits
injection withdrawal ratio, etc., then more accurate the outputs. Figure 5 is a graphical representation of a
predictions could be made. What is needed is a tool typical Neural Network.
that provides the professional the means to consider a
larger number of variables. Neural Networks are
ideally suited to this type of problem.
t Weight 11-H1
Input 1 \. Weight 11-H2
Hidden Node 1
• Weight H1..Q1
• Weight H1-02

Neural Network Primer • Weight 11·HM • Weight H1-00

A Neural Network, as defined by Dr. Robert Hecht- r Weight 12-H1

L
Nielsen (inventor of one of the first commercial
neurocomputers), is "a computing system made up of
.pm '1* :I'lltI2-H' 4++--f
Output
2

a number of simple, highly interconnected processing • Weight 12·HM

elements, which process information by its dynamic
state response to external inputs. ,,4 This definition
reflects the basic structure of a Neural Network as Output
that of a computer representation of a biological Hidden Node M o
neuron (nerve cell) that is interconnected with other
r Weight IN-H1 .Weight HM-01

neurons (a brain).5 Mathematically, Neural Networks l'Pm N \ W~1g1It IN-H' • W~ight HM·02

can be viewed as a multi-variable non-linear

'Ir Weight IN-HM • Weight HM-
regression. Neural Networks are not programmed to
find a solution, instead they learn by example. Where
most computer programs have the principles of Figure 5 - Typical Neural Network Architecture
chemistry and physics coded into the program itself as
equations, Neural Networks do not. Neural Networks The network is made up of a group of interconnected
do not have any a priori knowledge coded into the nodes. Biologically, the nodes represent neurons and
program at all. Conventional programs apply the the interconnections represent the synapse, resulting in
coded scientific principles in the program to the a simple brain. The largest network, biggest brain, the
problem's inputs to calculate an answer. A Neural author has built has 158 nodes and a total of 5400
Network learns an empirical relationship between the interconnections. In comparison, the human brain has
inputs and the desired outputs. The network is over 100 billion neurons and some of these neurons
presented with different examples of inputs and are connected to 10,000 other neurons.
outputs and relationships are learned after reviewing
the examples over and over again, as many as ten Every input and output is represented by an individual
million times. 6 It is capable of seeing what node. The nodes connecting the inputs with the
Professionals would see if they had the time and outputs are called hidden nodes. They are called
inclination to review the inputs and outputs a million hidden nodes because the end user of the network
times. Professionals who have reviewed inputs and does not know they exist. Every input node is
outputs over and over again and have learned a connected to every hidden node and every hidden
relationship have what is called years of experience. node is then connected to every output node. One
They are considered experts. name for this type of network is a "feed forward"

199
6 PREDICTING PRODUCTION USING A NEURAL NETWORK SPE 30202
(ARTIFICIAL INTELLIGENCE BEATS lIDMAN INTELLIGENCE)
network. It is called feed forward because the inputs The choice of which transform function and the
are presented to the network on one side and they number of hidden nodes is left to the Network
flow through the network to the outputs on the other designer, usually several combinations are tried and
side. the one yielding the smallest error is finally chosen.
The signals from the hidden nodes proceed in a similar
Every connection in a network is associated with a fashion to the output nodes where the final signal is
weight factor. These weight factors are independent scaled-up to yield the network's answer. The first pass
of each other. The mathematical workings of a of data through the network produces an answer that
network are simple. First the possible range in values is usually way off the mark since all the connection
for every input and output is determined. Then all weights are initially random numbers and have not yet
inputs and outputs are scaled for their range with 0 for learned anything. The networks answer is compared
the smallest possible value to 1 for the maximum to the correct answer and the network makes self-
possible value. At every node, any input to that node adjustments to all the connection weights to minimize
is multiplied by its weight factor and then summed this error.
with the product of the other inputs and their
This self-adjustment is the reason that another name
individual weight factors (see Figure 5). This resulting
for this type of network is IIback propagation ll
sum is processed by a transform function within the
network. The error is back propagated through the
node.
network by adjusting the connection weights in a
proportional fashion so the error is minimized. This
Two of the more commonly used transform functions whole process, feed forward and back propagate, is
are the hyperbolic tangent and logistic. They are repeated over and over again for every example the
IIsquashing functions;" that is, they pass along a signal network is supposed to learn. Learning is finished
with a value between -1 to 1 and 0 to 1, respectively. when the average error for all the examples is
This transform function provides non-linearity to the minimized. This iterative learning may take as many
network and constrains the node's signal within a fixed as several million passes of each example through the
range. Regardless of how lI excited ll an individual node network before an acceptable average error is reached.
gets, it has a fixed maximum signal that it passes on
through the network, just like their biological neuron Neural Networks are so good at learning that they can
counterparts. Figure 6 shows a single node and its use actually memorize all of the training examples. This
of the logistic function. usually produces a network that does not
IIunderstand, II it will be poor at applying its newfound
knowledge. It is just like the student who had a copy
of the answer sheet and just memorized the sequence
of answers. He may get a good grade on the test but
x 1 he does poorly when having to apply this knowledge.

(1+e-x J To keep the network honest requires an additional

data set of examples for testing the network. Every so
often the resulting best network to that point in
training time is used on these examples to see how
accurate the network is. As long as the network keeps
improving on this test set, training will continue.
When the network's error starts increasing on this test
Figure 6 Node's Transform Function set, then it is futile to continue training as the network
is just memorizing the answers.

200
SPE 30202 ROBERT J. BOOMER 7

The true test of the validity of a network is done by The professional typically looks only at the nearest
giving the network a "final exam." A third set of offset wells to a prospective drilling location. This
examples is used for this testing. The network error usually is the nearest four to eight wells. Since a
on this data set is the true measure of how good the Neural Network can easily handle more wells than
Neural Network performs. The selection of data for this, a Data Mask is utilized to capture data for wells
the Learning, Testing, and VerifYing data sets is not a that are not direct offset wells. The Data Mask is
matter to be taken lightly. To have a Neural Network shown in Figure 7. The Data Mask is a 5x5 grid
that you have confidence in requires that the data in covering 1,000 acres. Each square grid (1,320 feet
each of these three data sets be representative of the per side) equals 40 acres. This Data Mask is placed
problem you are trying to solve. In terms of Artificial over the drilling location's 40-acre tract and the data
Intelligence, the data must cover the Domain of the listed in Table 1 is collected for every well within the
problem. boundaries of the Data Mask. Since the number of
inputs in a Neural Network is a fixed number, the
It is not unusual to try various combinations of number of data values collected by the Data Mask
transform functions with a varying number of hidden have to be constant for every drilling location. This is
nodes and selected inputs before deciding upon the impossible as the number of wells within this Data
final network architecture. This entire process-feed Mask varied between less than 10 to greater than 50.
forward, back propagate, stop and test, and then The solution is to divide the Data Mask into three
continue training-is easily accomplished with concentric "rings." The "rings," labeled X, 10, and 9,
commercially available software. form a bullis-eye target with the drilling location
located at the center in the X-Ring. Data collected
Production Prediction Neural Network within each ring is then averaged so the data collected
in Table 1 is used with the Neural Network on a per-
The first step in building any Neural Network is to ring basis instead of a per-well basis. The point in
develop a data model that captures the data that a time that this current rate and cumulative data
professional would use to solve the problem. What represent is the calendar month preceding the month
inputs would an expert use to generate an output? that oil production for the newly drilled well is first
The review of the past methods of predicting reported to the state of New Mexico, i.e., initial
production is the starting point. Data used most often completion.
is either offset production data or reservoir
characterization data. Since production data is readily
available for all wells in the Vacuum Field and DATA MASK
reservoir characterization data is not readily available
for all the wells in the field, only production data is
. ------ ..... ----
selected as inputs. The specific production data
gathered for every offset well is shown in Table 1.

Current Rate Basis Cumulative Basis "DO'

Oil Oil
Water Water
Percent Oil Percent Oil
Injection Injection 1120'

Inj. - Withdrawal Ratio Inj. - Withdrawal Ratio I

Table 1 Collected Production Data Figure 7 Data Mask

201
8 PREDICTING PRODUCTION USING A NEURAL NETWORK SPE30202
(ARTIFICIAL INTELLIGENCE BEATS HUMAN INTELLIGENCE)
In summary, the data input items, data an expert might This is the verify data set, data the network had never
use to make a production forecast, for the Neural seen while training. It shows that the Neural Network
Network are the current rates and cumulatives for all slightly underestimates production 38 percent of the
the wells within a %-mile radius of any proposed time. It has a mean error of -19 BOPD and a mean
drilling location expressed as an average for three absolute error of 27 BOPD. This network is over 3.5
distinct locations from the drilling location as of the times (a mean absolute error of 27 BOPD for the
month before the drilling well is initially completed. Neural Network versus and a mean absolute error of
The outputs for the Neural Network are the one-, 99 BOPD for the Professionals) more accurate than
three-, six-, and twelve-month cumulative production the Professionals.
volumes for the proposed drilling location. All this
well data was collected for each of the 250 producing
wells completed in the Vacuum Field since 1970. To Initial, 3, 6, & 12 Month
gather the data for these 250 wells required that data Average Cumulative Oil (SOPD)
for 7883 offset wells be captured. A schematic of the T PREDICTED
network is shown as Figure 8. 1000 r--~-~-------~-....."...,

INPUT HIDDEN OUTPUT 100

NODES NODES NODES •
X-Rlng
10
•
OllRate

1st Month
i.,2~~~"-I~k2~~::::::;;:;~
1~---l--+-f ............t-H+-+--I---+--lH-H_--t----l-+--H-+-t+l
X-Rlng Cumulative
Cumulative -: OIL 1 10 100 1000
011 3rd Month
ACTUAL
X-Rlng
Water Rate
~~~~~~~§§;~~~~ Cumulative
011
6th Month
• Verify's -Perfect Predictor
Oumulatlve
011
41st Year Figure 9 Neural Network's Performance
Oumulatlve
9-Ring 011
Cumulative Conclusions
IIW

Statistical analyses of the networks' results are shown

Figure 8 Neural Network
in Table 2 along with the analysis of the Professionals'
results. As shown in Table 2, the Neural Network is
This network learned the relationship between all within ±20BOPD 51 percent of the time and within
these variables for about 164 randomly selected wells. ±lOBOPD 39 percent of the time. In a direct
An additional 74 randomly selected wells were used to comparison, the Neural Network out performed the
test the network to prevent memorization. The Professionals 93 percent ofthe time.
remaining 12 wells were used to verify the accuracy of
the Neural Network. Over 30 different network
configurations were tested before this network was Accuracy (% of the Time)
selected as the best. Training time took approximately PROFESSIONAL NEURAL
four hours on a 33mhz 486 PC. NETWORK
±20BOPD 8 51
Neural Network Prediction Accuracy ±10BOPD 5 39
Optimum 7 93
Figure 9 shows the same Confusion Matrix seen
earlier but with the Neural Network's predictions. Table 2 Performance Comparison

202
SPE 30202 ROBERT J. BOOMER 9

While this is quite an improvement, it still is not a

perfect predictor. If the Neural Network was VACUUM FIELD
presented with additional data that correlates with
production, then it would perform even better. An
obvious source of this relevant data is
geological/geophysical information. Just as the
Professionals used this information in making their
predictions, a Neural Network would also gain
knowledge if it had geological/geophysical data as an
input. Figure 10 shows the more familiar rate versus
time production plot for an average well. It also
shows the predicted profile for both the Professionals
and the Neural Network.

AVERAGE PRODUCTION PROFILE

'f SOPD
250 I I I i I ! i

200
i
······+..·
I
·1·· ·..··.. !..· ·..·.. t · I · ,.· ·
! i!
·'I'··..· ·r ·r I!
r····..·.·
150 1 j + .L + ++.
100 . · ··j···..· +· ··f ·I.. ··· ···· . ·..· ·.. j ·t· ·..·..t.. · .
50 ·..·..·!·· ··.. ·j· · i..· . · t·..· ····I ·..,..······ r.. .· i· ··.. ··t····..······[·..·..·..·..
o I 1 I ! i i II 1
1st Year
1 2 3 4 5 6 7 B 9 10 11 12 Average Produotion
Month tm50-100 BOPD ~100-150 BOPD _150-200 BOPD
-Actual .... Professional "'P- Neural

Figure 11 Neural Network Screening

Figure 10 Typical Performance
Summary
Another advantage, besides accuracy, the Neural
Network has over the Professionals is that it can A Neural Network that predicts production is a
quickly screen vast areas for those locations with the powerful tool that can reduce uncertainty in our
highest production potential, generating production predictions.
profiles for all possible drilling locations. Figure 11
shows the results of such a screening that covers Acknowledgments
several select properties, about eight square miles,
within the Vacuum Field. It would take a Professional The author thanks Texaco Exploration and Production
weeks or even months to generate these production Inc. for permission to publish this paper and to my
profiles while it takes only a couple of hours to build colleagues for their support during its preparation and
the data set and execute the Neural Network. review.

203
10 PREDICTING PRODUCTION USING A NEURAL NETWORK SPE 30202
(ARTIFICIAL INTELLIGENCE BEATS HUMAN INTELLIGENCE)
References

1Geologic and Engineering Approaches in

Evaluation ofSan AndreslGrayburg Hydrocarbon
Reservoirs-Permian Basin, D. G. Bebout and P. M.
Harris (ed.), Bureau ofEconomic Geology, Austin,
Texas (1990) iii

2Purves, William J.: "Re'servoir Description of the

Mobil Oil Bridges State Leases (Upper San Andres
Reservoir), Vacuum Field, Lea County, New Mexico,"
Geologic and Engineering Approaches in Evaluation
ofSan AndreslGrayburg Hydrocarbon Reservoirs-
Permian Basin, D. G. Bebout and P. M. Harris (ed.),
Bureau ofEconomic Geology, Austin, Texas (1990)
87-112

3Klimasauskas, Casimir C.: "Applying Neural

Networks: Part III,: PC A/May/June 1991 53,20-24
4Caudill, Maureen: "Neural Networks Primer Part I,"
A/ Expert December 1987212,46-52

5Klimasauskas, Casimir C.: "Applying Neural

Networks: Part I," PC A/ JanuarylFebruary 1991 5 1,
30-33

6Wasserman, Phillip D. :Neural Computing Theory

and Practice, Van Nostrand Reinhold, New York
(1989) 1-26

204