Applications of Neural Network in Manufacturing

advertisement
Proceedings of the 29th Annual Hawaii International
Conference on System Sciences -
1996
Applications of Neural Network in Manufacturing
Ramesh Rajagopalan,Ph.D. * and Purnima Rajagopalan+
* Industrial and SystemsEngineering, Mercer University, Macon, GA 3 1207
rajr@egr.mercer.peachnet.edu
+ M IS Manager, Galaxy Services Group, Macon, GA 3 1210
Abstract
Neural network is a model of brains’s cognitive process.
Neural network originated as a model of how the brain
works. Neural network research has its beginnings in
psychology. Today neural network methods are being
used to solve numerous problems associated with
manufacturing operations. A review of neural network
applications to problems in production and operations
management is presented. Applications reviewed in this
paper include character, image andpattern recognition,
managerial decision making, manufacturing cell design,
tool condition monitoring, real-time robot scheduling and
statistical process control. Methods and structures of
neural network are explained.
I. Introduction to neural network
Over the past several years, various methods based on
such areas as operations research, statistics, computer
simulation, control theory have been developed and
applied to solve a wide spectrum of problems in
manufacturing. Today’s manufacturing environment is
characterized
by
complexity,
inter-disciplinary
manufacturing functions and ever growing demand for
new tools and techniques to solve difficult problems.
Neural network offers a new and intelligent alternative to
investigate and analyze challenging issues related to
manufacturing. In this section, an introduction to neural
network and two commonly used neural network methods
will be provided.
In supervised learning, a set of data, called a training
data set, is used to help the network in arriving at the
appropriate weights [40]. A teacher teaches the network
and gives results of the output corresponding to the input.
The inputs as well as side information indicating the
correct outputs are presented to the network[20].
The
network is also ‘programmed’ to know the procedure to be
applied to adjust the weights and thus the network has the
means to determine whether or not its output was correct
and the means to apply the learning law to adjust its
weights in response to the resulting errors[20].
Weights are generally modified on the basis of the
errors between desired and actual outputs in an iterative
fashion and one of the widely used training algorithms is
the “Delta Rule” [20]. The neural network learns the
desired outputs by adjusting its internal connection
weights to minimize the discrepancy between the actual
outputs of the system and the desired outputs [32].
Neural network is used to capture the general
relationship between variables of a system that are
difficult to analytically relate. Neural network has been
described as “brain metaphor of information processing”
or as “a biologically inspired statistical too1”[35]. It has
the capability to learn or to be trained about a particular
task, its computational capabilities and the ability to
formulate abstractions and generalizations.
Neural
network has an organization similar to that of a human
brain and it is a network made up of processing elements
1060-3425/96 $5.00 0 1996 IEEE
called neurons. Neurons get data from the surrounding
neurons, perform some computations, pass the results to
other neurons. Connections between the neurons have
weight associated with them. In neural network, the
knowledge is stored in the network’s interconnection
weights in an implicit manner, learning takes place within
the system and plays the most important role in the
construction of an neural network system. The neural
network system learns by determining the interconnection
weights from a set of given data [40]. Learning in neural
network can be supervised, unsupervised or based on a
combined unsupervised-supervised training [20].
In unsupervised learning a neural network operates in
self-organization mode. In this self-organization mode, a
competition mechanism is used to select processing
elements due to which their weights are modified. The
first part of the competition involves competition between
the simple-elements that he within same sample. Winners
from these competitions compete with other processing
elements in their layer that have won similar competitions
in order to be single processing element on their layer
447
Proceedings of the 1996 Hawaii International Conference on System Sciences (HICSS-29)
1060-3425/96 $10.00 © 1996 IEEE
Proceedings of the 29th Annual Hawaii International
III. Neural network applications
which has the weight changed during training. In selforganized training, the network is given many input
characters and no information is given to the network as to
what each example corresponds to. There are many layers
and each layer corresponds to one character more than the
other layers. The importance of unsupervised learning is
that the system does not need to know the correct answer
in order to solve a problem. The system learns a pattern
from repeated exposure to it and is able to recall the
learned pattern when it solves a categorization or pattern
matching problem [40].
Neural network offers a method for incorporating and
processing qualitative knowledge and have the additional
advantage of formalizing machine learning in an explicit
manner, it aspires to imitate human intelligence in its
totality [40]. Neural network has been applied to a wide
variety of problems, ranging from traveling salesman
optimization to vision problems. Neural network is best at
solving classification problems and it also has the added
advantage of performing successfully where other
methods often fail - recognizing and matching
complicated, vague, or incomplete patterns [40].
Neural network can also employ a hybrid approach in
which learning is based on combined unsupervisedsupervised learning. The hybrid approach first uses
unsupervised learning to form clusters and the labels are
then assigned to the clusters identified and a supervised
training follows[20].
Neural network is used to convert text to speech, for
natural language processing, for example, for deriving
language rules, recognition of characters and handwriting,
image processing and pattern recognition. This line of
research has immediate use in banking, credit card
processing, and other financial services where reading and
recognizing handwriting on documents is crucial [40].
II. Neural network methods
Two methods are commonly used in neural network
applications - neocognitron and back propagation.
Neocognitron is a hierarchical network made up of many
layers and it’s organization is like that of the visual cortex.
Neocognitron is a method of pattern recognition. It has the
capability to recognize shapes and sizes of characters even
if it involves noise and distortion. Back propagation
allows the training ofmulti-layer networks, it is a powerful
and practical tool for solving problems that would be quite
difficult using conventional computer science techniques
and these problems range horn image processing to speech
recognition to character recognition forecasting to
optimization[5].
Due to rapid development of neural network methods
and tools, neural network has generated tremendous
amount of interest solving manufacturing
related
problems. Neural network is used to address issues
relating to manufacturing process planning and process
control, manufacturing system design, operational decision
making, and resource scheduling. The Table 1 provides
a summary of various neural network applications in the
general and manufacturing categories. The remainder of
this section is organized into two subsections. Section III. 1
describes general applications of neural network in pattern
and character recognition. Applications of neural network
to manufacturing problems are presented in section III.2.
Back propagation is the best-known supervised
learning method for Neural Network with three or more
layers[40].
In the back-propagation procedure, the
network fast uses the input data set to produce its own
output, and then compares this with the desired output, if
there is no difference, no learning takes place else the
weights are changed to reduce the error term [lo]. One of
the most important steps in the development of a neural
network is the development of the training data set and the
training data set must be constructed carefully to include
examples of many different operational scenarios.
Approximately
equal numbers of examples of each
In the
scenario should be included in the data set.
following section applications of neural network to a wide
variety of manufacturing planning, design and control
problems are presented.
III.1 General applications
A character recognition problem using neocognitron to
translate alphabets in one language to alphabets in another
language is the described by Sankamarayanan [ 3 11. The
author used the method of neocognitron to translate
English letters to Greek letters. The method is quite
versatile and may find widespread application as a general
pattern-recognition paradigm. A software system using
PASCAL was developed [31]. The software system
implements a neocognitron based self-learning neural
network. The input is a character with low-resolution.
Neocognitron is not affected by scale change, shift in
position, noise and distortion. It shares the weight during
training. This weight sharing method is used so that
whenever the weight of an element in a layer is modified,
448
Proceedings of the 1996 Hawaii International Conference on System Sciences (HICSS-29)
1060-3425/96 $10.00 © 1996 IEEE
Conference on System Sciences - 1996
Proceedings of the 29th Annual Hawaii International
Conference on System Sciences -
1996
through a visual system and recognized correctly by the
multilayered hierarchical network using self-organization.
Table 1. Summary of neural network applications.
To recognize handwritten digits using a machine and
using pattern recognition is difficult.
This is of some
interest to recognize ZIP codes in addresses written by
hand in mail. Neural network and its computational
properties have attracted the interest of researchers in the
area of machine perception by presenting an exciting,
complementary alternative to symbolic processing
paradigms[29]. The neural network gives results quickly
using unsupervised training. There are many different
neural network models.
Fukushirna et a1.[13j
The method of neocognitron is used to recognize
handwritten alphanumeric characters. This neural network
can acquire the ability to recognize patterns by learning,
and can be trained to recognize any set of patterns, it also
has a large power of generalization, presentation of only a
few typical examples of deformed patterns is enough for
the learning [13].
The neural network not only makes use of classic
notions of dynamic learning systems, but also has at its
core, the idea that a system may use relatively static sets of
previously learned information analogous to what
cognitive psychologists might refer to as schemas or
strategies[25]. In particular applications, such as optical
character recognition, it is possible to train a network by
presenting it with the finite universe of prototypical sets of
pixels that may lead to a correct solution most of the
time[25].
When the training is completed the neural
network is found to be stationary. Rarely when the neural
network is static, it fails to recognize the input figure and
in that case the dynamic neural network system is used.
These two networks are combined in an image recognition
unit that has the capability to exceed the expected
Neural
performance of human beings at this task[25].
network is used in the image recognition and pattern
recognition methods. The neural network model should be
synthesized in order to endow it with pattern recognition
capability like that of a human being. A cell in a deeper
layer generally has a tendency to respond selectively to a
more complicated feature of the stimulus patterns and, at
the same time, has a larger receptive field and is less
sensitive to shifts in position of the stimulus patterns, SO
each cell of the deepest layer of the network responds
selectively to a specific stimulus pattern and is not affected
by the distortion in shape or the shift in position of the
pattem[l4].
all other processing elements in the layer immediately
adopt new weights. The input to the network is in the
form of zeroes and ones (binary). Tests are conducted for
many iterations. The goal of the network is to identify
accurately character present in the input image. At each
layer of network input image is analyzed within the
surrounding area that becomes bigger and bigger till the
entire input character is classified. Several examples of
translation of English letters such as “a” and “b” to
corresponding Greek letters such as “~1”and “p” are given
in [31].
Neural network using neocognitron can recognize
handwritten numerals of various styles of penmanship
correctly, even if they are considerably distorted in
shape[31]. Although the author shows results for the
recognition of alphabet, the neocognitron can be trained to
recognize other set of patterns such as Arabic numerals,
geometrical shapes, or others. A pattern can be seen
III.2 Manufacturing applications
Moon and Chi [27] have utilized the generalization
449
Proceedings of the 1996 Hawaii International Conference on System Sciences (HICSS-29)
1060-3425/96 $10.00 © 1996 IEEE
Proceedings of the 29th Annual Hawaii International
capability of neural network models to solve the part
family formation problem. The approach presented by the
authors, combines the useful capacities of the neural
network technique with the flexibility of the similarity
coefficient method [27]. Manufacturing information, such
as the sequence of operations, lot size, multiple process
plans were given special consideration in their approach to
solve generalized part family formation problem. The
authors also point out that the method is flexible and can
be efficiently
integrated with other manufacturing
functions. The authors conclude that neural network can
address the Group Technology (GT) family formation
problem efficiently.
1996
network is started by providing an output signal to the
chosen neuron. A threshold value, which determines the
number of clusters as well as the degree of similarity
within a cluster is then decided. Neurons with activation
values greater than the threshold value are grouped
together. Thus part families and machine groups are
identified. The method then chooses a neuron arbitrarily
which has not been assigned to any family or group.
These steps are iteratively carried out until all the neurons
are clustered.
Wu describes two applications of neural networks to
solve manufacturing problems - to form group technology
based manufacturing cells and to monitor cutting tool
condition.
Kaparthi and Suresh [20] present a neural network
clustering method for the part-machine grouping problem.
This method is based on a neural network algorithm to
support procedures like production flow analysis. A
neural network clustering algorithm using similarity
coefficients, is used to solve the part-machine grouping
problem.
The neural network method is based on
unsupervised learning. In the part-machine matrix, each
row (part) is considered as a vector in a higher
dimensional space and every dimension corresponds to a
machine type, and the number of dimensions is given by
the total number of machine types required for all parts
[20]. The authors have shown that the neural network is
capable of handling large data sets.
To form a manufacturing cell, a neural network was
made to operate on the classical machine/components
matrix Patterns representing respectively the machine
characteristics and the component characteristics are used
as input to the network. This approach was tested on two
such matrices, one consisting of 10 jobs and 15 machines,
and the other 14 machines and 24 parts. The author
concluded that the results produced by the neural network
approach compare well with the classical clustering
techniques [ 3 71.
The monitoring of tool condition during metal cutting
operations is a classical production topic for which a
substantial amount of research work has been carried out,
the problem is to monitor accurately the condition of the
cutting tool on-line so that a worn-out tool can be detected
quickly and subsequently replaced [37]. The methods to
detect tool failure are based on the level of cutting force,
vibration or acoustic emission, and these are measured by
sensors that are attached to the tool. But the main
difficulty is how to detect the signal pattern produced by
a worn tool.
Lee et al., [24] present a method for part family
formation, machine cell identification, bottleneck machine
detection and the natural cluster generation is done using
a self-organizing neural network. The authors argue that
the generalization ability of the neural network makes it
possible to assign the new parts to the existing machine
cells without repeating the entire computational process.
The authors show that neural networks can learn from a
given set of patterns and are able to generalize this
knowledge to other similar problems. This property
makes them useful in small and medium-size batchmanufacturing systems where training data are limited and
new parts are continuously encountered. The authors
point out that their method based on neural networks is not
significantly influenced by the size of the machine-part
matrix and hence it is appropriate for solving large-scale
industrial problems.
Some of the signal components are selected from a
number of raw input data which are more sensitive to tool
ware, but less sensitive to process noise. A set of training
inputs are chosen which consists of signal patterns
representative of both fresh tool cutting, and worn tool
cutting and these are used to train a single hidden layer
back-propagation neural network with a 8-3-l structure,
i.e. 8 input processing units, 3 hidden processing units and
1 output processing unit, processing units are the building
bricks of neural network [37]. Target outputs of 0.01 and
0.99 aTe set for fresh and worn cutting patterns.
Moon [26] presents a neuro-computing model to
Similarity
identify
part family/machine
groups.
coefficients based on, for example McAuley’s method, is
The similarity coefficients are used as
determined.
connection weight values for neurons within the machine
and part layer. A neuron from either the part layer of the
machine layer is then chosen arbitrarily.
The neural
After training, each layer of the neural network acts as
a signal filter so that the neural network suppresses noise
and increases the signal/noise ratio step by step as the input
450
Proceedings of the 1996 Hawaii International Conference on System Sciences (HICSS-29)
1060-3425/96 $10.00 © 1996 IEEE
Conference on System Sciences -
Proceedings of the 29th Annual Hawaii International
patterns propagate through the network, the trained neural
network produced a 94% success rate for worn pattern
detection [37]. Neural network monitors successfully the
tool condition in metal-cutting operations.
Conference on System Sciences -
1996
The relative importance of various manufacturing
decision making criteria and the overall performance of a
manufacturing system is impossible to analytically
estabhsh[ll]. Chryssolouris et al., [l l] explore the use of
neural networks for identifying the relative importance of
decision criteria. Simulation and a neural network are
used to establish adequate weights of the criteria for the
decision-making process at the work center level. A
procedure for determining operational policies for
manufacturing systems has been presented. The authors
use simulation results to train a neural network which then
prescribes an operational policy suitable for achieving a set
of goal performance measures [ 111. The proposed neural
network procedure determines suitable criteria weights for
an entire sequence of multiple-criteria decisions, The
authors conclude that the method is better suited to
complex applications involving chains of decisions, such
as job shop scheduling, while the conventional methods
are likely better suited to isolated, single decisions [ 111.
Yih et al.,[39] uses a hybrid method that combines
human intelligence, optimization and neural network to
solve a real-time robot scheduling problem in a circuit
board production line. The hybrid method is divided into
three phases. ln phase I, a simulator collects data from
human schedulers. Phase II constructs Semi-Markov
decision models which are then used to fmd solutions to
the scheduling problem. Solutions derived from phase II
are then fed as training data for the neural network. In the
neural network model, there exists three layers: an input
layer, hidden layer, and output layer. The nodes between
the input layer and hidden layer are fully connected, as are
those between the hidden layer and the output layer. The
input layer contains nine nodes, each correspond to one
attribute in the state definition and the output layer has six
nodes that represent six possible decisions in the system.
For the hidden layer twenty nodes are selected after
systematically experimenting with various hidden nodes
based on the performance of the resulting neural network.
The neural network was trained for twenty hours on a 486
IBM-compatible personal computer. The duration of
training was determined by observing the level of errors of
the neural network and the decrement in error rate.
Training the neural network was concluded when the error
levels were low and decrement in error rate was
approaching zero. The resulting network model was then
evaluated in the real-time robot scheduling problem [39].
Yao et a1.,[38] uses a distributed neural network of
coupled oscillators to solve an industrial pattern
recognition problem. The problem addressed is machine
recognition of industrial screws, bolts, etc. in simulated
real time in accordance with tolerated deviations from
manufacturing specifications. Inputs are preprocessed and
represented as 1 x 64 binary vectors. The supervised
neural network uses the backpropagation method to
accomplish the pattern recognition task. The authors state
that the neural network method performs better than a
standard Bayesian statistical method.
IV. Summary
Cook and Shannon [lo] present a methodology to
predict the occurrence of out-of-control process conditions
in a composite board manufacturing facility. This method
is developed using neural network theory. The neural
network,
using
back-propagation
method,
was
successfully trained to represent the process parameters.
The trained neural network was able to successfully
predict the state of control of the specific manufacturing
process parameters with 70% accuracy. The learning rule
used in this research was the generalized delta rule which
is an error-correcting rule that has been used in various
applications including converting printed text to speech,
controlling
robot arms, and selecting good loan
The back-propagation learning
applications [lo].
algorithm uses a gradient- or steepest-descent heuristic that
helps a network to organize itself in ways that improve its
Continuous manufacturing
performance over time.
processes with measurable process parameters are
promising application areas for neural network modeling
The methods, structures and applications of neural
networks are presented in this paper. Ever since the
beginnings of research in artificial intelligence, neural
network has shown tremendous potential in solving
complex problems. Neural network methods based on
supervised and unsupervised learning are used in
numerous manufacturing applications. Solution methods
based on neural network are elegant and efficient in
solving a wide range of problems in manufacturing that
are critical for companies’ survival in today’s competitive
environment. Applications of neural network presented in
this paper are only a small sample of a large number of
potential applications.
BIBLIOGRAPHY
VI.
[lOI.
451
Proceedings of the 1996 Hawaii International Conference on System Sciences (HICSS-29)
1060-3425/96 $10.00 © 1996 IEEE
Abutaleb, A.S., “A Neural Network for estimation of
forces acting on Radar Targets”, Neural Networks,
vo1.4, 1991, pp.667-678.
Proceedings of the 29th Annual Hawaii International
PI.
Atiya, F. A., “An Unsupervised Learning Technique
for Artificial Neural Networks,” Neural Networks,
vo1.3, 1990, pp. 707-711.
[33.
M.
Conference on System Sciences -
1996
D51.
Fukushima, K., and Miyake, S., “Neocognitron: A
New Algorithm for Pattern Recognition Tolerant of
Deformations and Shifts in Position,” Pattern
Recognition 15(6), 1982, pp. 455-469.
Baran, N., “The Outlook for Pen Computing,” BYTE,
September 1992, The Worldwide
Computing
Authority.
V61.
Fukushima, K., “Cognitron:
A Self-organizing
Multilayered Neural Network,” Biological Cybernetics
20, 1975, pp. 121-136.
Bebis, G.N., and Papadourakis, G.M., “Object
Recognition
using invariant object boundary
representations and Neural Network Models,” Pattern
Recognition, Vo1.25, No.1, 1992, pp.25-44.
P71.
Jacobs, R.A., “Increased Rates of Convergence
through Learning Rate Adaption,” Neural Networks,
~01.1, 1988, pp.295-307.
r51.
Blum, A., Neural Networks in C++ - An ObiectOriented Framework for Building Connectionist
Systems, John Wiley and Sons, New York, NY, 1992,
pp.l-10.
U81.
Johnson, K., Daniell, C., and Burman, J., “Feature
Extraction in the Neocognitron,” IEEE International
Conference on Neural Networkq Vol. 2, 1988, pp.
117-126.
161.
Carpenter, G.A., and Grossberg, S., “The ART of
adaptive pattern recognition by a self-organizing
neural network,” Computer, 21 (3), 1988, pp.77-88.
D91.
Kaparthi, S., and Sure&, N.C., “A neural network
system for shape-based classification and coding of
rotational parts,” International Journal ofProduction
Research, 29 (9), 1991,pp.1771-1784.
[71.
Caudill, M., “Back Propagation Networks,” Naturallv
Intelligent Svstems, Chapter 14, 1990.
PO].
Kaparthi, S., and Suresh, N.C., “Machine-component
Cell Formation in Group Technology: A Neural
Network
Approach,” International
Journal of
Production Research 1992, Vol. 30, No. 6, pp. 13531367.
Pll.
Khotanzad, A. and Lu, J.H., “Distortion Invariant
Character Recognition by a Multi-layer Perceptron and
Back-Propagation Learning,” IEEE International
Conference on Neural Networks 1988 - Volume I,
pp.625632.
P21.
Lippmann, R.P., “Pattern classification using neural
networks,”
IEEE
Communications
Magan’ne,
November, 1989, pp.47-64.
~231.
Lippmann, R.P., “An introduction to computing with
nemal nets,”IEEE A S S P Magazine, April, 1987, pp.422.
PI.
Caudill, M., “Neural Networks, Primer Part II,” AI
Expert, February, 1988, pp. 55-61.
PI.
Caudill, M., and Butler, C., “A Neural Network Model
for Selective Attention,” Proceedings of the IEEE
First International Conference on Neural Networks.
Vol. 2, 1987, pp.ll-18.
uw
[ill.
Cook D.F., and Shannon, R.E., “A predictive neural
network modelling system for manufacturing process
parameters,” International Journal of Production
Research, 1992, Vol. 30, No.7, pp.1537-1550.
Chryssolouris, G., Lee, M., Domroese, M., “The Use
of Neural Networks in determining Operational
Policies for Manufacturing Systems,” Journal of
ManufactutingSystems, Volume lOiNo.2, pp.166175.
WI.
Eberts, R., “Using Neural Nets for Assistance in
Interacting with Computerized Systems,” Proceedings
ofthe I992 NSF Design and Manufacturing Systems,
pp. 687-690.
~241.
Lee, H., Malave, CO., and Ramachandran, S., “A selforganizing neural network approach for the design of
cellular manufacturing systems,” Journal ofIntelligent
Manufactuting (1992)3,325-332.
P31.
Fukushima, K., and Wake, N., “Handwritten
Alphanumeric
Character Recognition
by the
IEEE Transactions on Neural
Neocognitron,”
Networks. Vol. 2,1991, pp. 355-365.
~251.
Mehr, D., and Richfield, S., “Neural Net Application
to Optical Character Recognition,” IEEE Zntemafional
Conference on Neural Networks, Vol. 4, 1987, pp,
771-776.
[141.
Fukushima,
K., Miyake,
S., and Ito, T.,
“Neocognitron:
A Neural Network Model for a
Mechanism of Visual Pattern Recognition,” IEEE.
Trans. Systems,Man, and Cybernetics, Vol. SMC-13,
No. 5, 1983, pp. 826-834.
[261.
Moon, Y.B., “Establishment of a neurocomputing
model for part family/machine group identification,”
JoumalofIntelligentManufacturing,
(1992)3, pp.173182.
452
Proceedings of the 1996 Hawaii International Conference on System Sciences (HICSS-29)
1060-3425/96 $10.00 © 1996 IEEE
Proceedings
of the 29th Annual Hawaii
International
v71.
Moon, Y.B. and Chi, SC.,” Generalized Part Family
Formation Using Neural Network Techniques,”
Journal ofManufacturing Systems, Volume 1 l/No. 3,
pp.149-159.
P81.
Pao, Y., Adautive Pattern Recoenition and Neural
Networks, Addison-Wesley, New York 1989.
P91.
Pawlicki, T. F., Lee, D-F., Hull, J.J., and Srihari, S.N.,
“Neural Network Models and their Application to
Handwritten Digit Recognition,” IEEE International
Conference on Neural Networks, Vol. 2., 1988, pp. 6310.
[301.
Rumelhart, D.E., Hinton, G.E., Williams, R.J.,
“Learning
Internal Representations by Error
Propagation,” Parallel Distributed Processing, Chapter
8, vol. 1, 1986.
[311.
Sankamarayanan,
P.,
“Application
of
the
Neocognitron to the Character Recognition Problem,”
unpublished Master’s Thesis, Central Michigan
University, December 1992.
~321.
Srinivasa, N., Kalle, P., Ziegert, J.C., and Smith, S.,
“Neural Network Prediction of Machine Tool Error
Maps,” Proceedings of the I992 National Science
Foundation
(NSF) Design and Manufactun’ng
Systems, pp. 1099-1103.
[331.
Vogl, T.P., et al., “Accelerating the Convergence of the
Back-Propagation Method,” Biological Cybernetics,
vol. 59,198X, pp.257-263.
[341.
Wasserman, &Ural Computing, pp. 167-188.
[351.
Wilson, R., and Sharda, R., “Neural Networks,”
OR/MS TODAY, A Joint Publication of ORSA and
TIMS, August 1992, pp. 36-42.
[361.
Winograd, T., and Flores,
Comnuters and Coenition.
[371.
Wu, B., ” An introduction to neural networks and their
applications in manufacturing,” Journal of Intelligent
Manufacturing (1992)3, pp.391-403.
1381.
Yao, Y., Freeman, W.J., Burke, B., and Yang, Q.,
“Pattern Recognition by a Distributed Neural Network:
An Industrial Application,” Neural Networks, Vol. 4,
1991, pp.103-121.
L391.
Yih, Y., Liang, T-P, and Moskowitz, H., “Robot
Scheduling in a Circuit Board Production Line: A
Hybrid OR/ANN Approach,” ZZE Transactions,
Volume 25, Number 2, March 1993, pp.26-33.
F.,
[401.
m
453
Proceedings of the 1996 Hawaii International Conference on System Sciences (HICSS-29)
1060-3425/96 $10.00 © 1996 IEEE
Conference
on System Sciences -
1996
Zahedi, F., “An Introduction to Neural Networks and
a Comparison with Artificial Intelligence and Expert
Systems,“Znterfnces 21: 2 March -April 1991, pp.2538.
Download