Artificial Neural Network Management System

Gao Ningning
Information Engineering College
Central University for Nationalities
Beijing 100081, China
[email protected]

Abstract—This paper focuses imitation of a biological neural technology, is undoubtedly promoting the applied research on
network and its applications. From the BP and the RBF network neural networks. This system reflects the wisdom and
technologies, it is proposed a schema of bionic neural network experience of mankind, and widens the recognition channels of
model based management system (ANN MS), pointed that the the objective world.
system component level (model) development of practical value,
and discussed the basic applications of the system: modeling
classification, model mining, the detection and diagnoses, and so II. MAIN TECHNOLOGY AND IMPLEMENTATION
on.
A. BP Network and Its Learning Algorithm
Keywords- bionic neural network; network management system; The output signal of multilayer feed forward neural
component development; modeling classification;model mining network is the complete responses to the network input layer
excitation pattern on source nodes, that is, the input signal
I. RESEARCH BACKGROUND imports from the input layer, passes through the hidden layer
reaching output layer. The output signal is obtained from the
Since the mid-80s, once quiet ANN (Artificial Neural
output layer as Fig.1 shown.
Networks, ANN), has been re-launched a new artificial neural
network craze being a huge response to the two studies on Be noted that the hidden layer can be single or multi-layer;
artificial neural networks from the U.S. Academic Science. signal can be transmitted between nodes only through transfer
function. Here the transfer function is selected as the S-type,
Artificial neural network is formed by the artificial neurons
that is,
(simulation of biological neuron feature structures) within the
inter union. ANN artificial neural network can be used to either
produce a comprehensible biological neural network, or solve
the problem of artificial intelligence without having to create a
true model of biological system As for the definition of an From Fig.1 it is known that work signal is being transmitted
artificial neural network, there are various different opinions. forward, in which the connecting weight doesn’t change. If the
output value is less than expected value, then an error signal
Currently the most widely used definition is T. Koholen’s turns to transferred back propagation. Process of back
that “an artificial neural network is a network that is organized propagation of error signal is called learning, during which
by simple units with extensive and parallel interconnection. Its weights are adjusted by the error feedback.
organization can simulate the biological neural system to make
interactive responses with real-world objects.” According to learning rules, there is:
Despite the neural network theory and its application have
made great progress since 1980s, people have a long distance
to reach understanding and researching the biological nervous
system, such as central nervous system and so on. Furthermore,
how to understand an artificial neural network? How to use it
and explain its results? Admittedly, an artificial neural network
shows powerful information processing capabilities, every time
when today's computers are hard in dealing with the nonlinear
characteristic problem. Although such complexity can be
resolved by a neural network, in practice, neural networks can
not lonely be used for the solution problem, corresponding
system, even a corresponding set of management system, must
be developed.
Multilayer feed forward neural networks are relatively
mature both in theory and in applications. The artificial neural
Figure 1 Schema of Feed Forward Neural Network Model
network management system (ANN MS) developed by such

978-1-4244- 7618-3 /10/$26.00 ©2010 IEEE

which the new weight from neuron j to neuron i equals to the layer and output layer, the difference for expected value T and
old value adding the synaptic weight adjustments. The last output Y of hidden layer should be the product factor of the
component can be expressed as orthogonal matrix . The iterative formula of gradient method
for adjusting weight matrix is as follows

among which Ș is learning step (or learning rate parameter), e

(m) is the error between the expected output and the actual Artificial neural network (ANN) is still cutting-edge
output, and m is the number of iterations. technology by now. Its applications are growing continuously,
which mainly include function approximation, modeling, time
Forward neural network input-output relationship can be sequence analysis, pattern recognition and control etc. BP
seen as nonlinear relationship; and forward neural network uses networks and RBF networks all belong to the multilayer feed
error backward transmitting (Error Back Propagation) learning forward artificial neural network. Compared with BP network,
algorithm, so called BP networks. BP algorithm in practice RBF network has a faster learning ability, greater
often appears slow network convergence phenomenon, relating approximation ability, and better classification ability.
with size selection of learning step: large Ș value makes
networks converge quickly, but too large to be stable; while The current study for ANN network is limited by: results of
small Ș value makes networks to avoid instability, but the limited brain research; the lack of a complete, mature
convergence rate slow. theoretical system; and the style of a strong strategy and
experience. These will be broken or transformed as progressing
To emit these potential conflicts, rewrite formula (3) in the in science, and establishment of a new theory.
variable t of time sequence, t: 0 to m:
III. ANN MS SYSTEM SOLUTION
The identification of BP / RBF neural networks is reflected
among which ¢ is the momentum constant, when 0 İ | ¢ | in the nonlinear I / O modeling, and the corresponding
<1, the sequence converges. algorithm is more perfect; so it is time for them to be out of the
theoretical research, and to enter applied research. This is just
Performance of a BP network is approximating any the significance of developing an Artificial Neural Networks
nonlinear relationship with good generalization ability, strong Management System (ANN MS), further more, the profound
fault tolerance, slow convergence of the learning procedures, significance is finding the limitations of ANN network,
and easily trapped in local minimum, and so on. developing a more anthropomorphic “brain” through applied
research.
B. RBF Network and Its Learning Algorithm
ANN MS system is structured artificial neural network
Radial Basis Function (RBF) network, referred as RBF model from the view of bionics, and its application is based on
network, is a three-layer forward neural network, that is, input the model, therefore, its main functions are to establish and
layer, hidden layer with radial basis function neurons and simulate BP network, to establish and simulate RBF networks,
output layer with linear neurons. Hidden layer output to research the application model basically, and so on. The
expression is: system is a management system based on BP / RBF artificial
neural network theory and application research.

If Gaussian function is chosen as the radial basis function Fig2 is an ANN MS system function block diagram. It
radbas, then: shows that the total function of the artificial neural network

where z = ý w - x ý • b.
Therefore, the RBF learning algorithm can be determined
by obtaining the input layer and hidden layer weights , as
well as hidden layer and output layer weights respectively,
like the following:
Using unsupervised k-means clustering method to learn the
sample, determine base function center ( ) and radius ( : may
also be called a Gaussian variance) from the sample data, to
obtain connection weights and the basis function y.
Using method of supervised learning, train samples in order
to obtain the adjustment of connection weights : using the
negative gradient method which Gaussian function of the
orthogonal matrix can be obtained from the gradient
characters; being linear of transfer function between hidden Figure 2 ANN MS system function model

Identify applicable sponsor/s here. (sponsors)

management system has five major components: System development of science, it is undoubtedly a bright future. Feed-
Management, BP Networks Creation and Simulation, RBF forward neural network is a sort of ANN networks in early
Networks Creation and Simulation, Basic Applications of study: inspired by each source node in input layer, the first
Models, and Help, etc. To be clear, the same sub-function hidden layer receives input signal, the output of it is as the
blocks of creation and simulation for both BP network and input of second hidden layer, then the output of it is as the third
RBF network are drawn together, so the diagram is more hidden layer input, and so on. The network output layer is the
clearly. From the diagram, the system has a component network response to input excitation. For ordinary users, the
function, which means the system can be used to do the hidden layer neurons are just like a black box, and their
development of component level, or it can be said neural
energies will be used to the data preparation, experience, or
network will be applied as a model.
undergoing, as well as plan and break strategies, etc., as in [6]
As mentioned earlier, neural networks are applied as a study, 7 kinds of similar comparison features are as inputs for a
functional model, which requires models to be managed. Only similarity detector.
object-oriented database technology, can meet the
In ANN MS system, networks belong to common feed
requirements. Object-oriented data model considers the
forward neural networks. BP network that is multilayer
combination of data describing objects and their behavior
preceptor, the learning algorithm (BP algorithm) and its
Object-oriented database management is mainly used to
expand the objects, long-effect on transaction processing and improved algorithm are more perfect in theory. There are many
pattern evolution, etc. Object-oriented database interface is examples of successful application. On the BP base, three-tier
mainly used to support message passing, to provide RBF network is proposed, which is the common result of BP
comprehensive computer database language. Object-oriented network and the results of biological research. Mapping
database technology is composed of the data model, database function of RBF neurons in single hidden layer is usually
management and database interface. In addition, to realize chosen Gaussian function, based on results of more recent
basic application functions it should be prepared for data biological research: that is, different parts of the brain cells
mining database, in which data, algorithms, and mining ways reflect the intensity of external stimulation differently. The
are stored. common features of the two networks are massive and parallel
operations, processing and storing simultaneously, as well as
In fig2, function of BP network creation and simulation
adaptively self learning, and so on.
has sub-function ---- "to improve the network", which can
adjust learning step with the help of a wizard. In general, Although the BP/RBF network has different characteristics
reducing the learning step, increasing the momentum factor all of good applications in many areas, it is essentially based on
can speed up the convergence, and the organic combination of the approximation capability of the network Through the
the two can make the network convergence reach the best state. study; the network has obtained the system function with the
For any complex, nonlinear, and never touched problem, you tolerant error under the guidance of teacher signal. These are
can start from the BP network to learn, or to study, while for enough to build static and bionic systems, that is, under the
the network with only single hidden layer, RBF network is prior knowledge (teacher signal), the established model
reasonable to use. When researching, pay particularly attention
(ANN), is formed a subsystem based on the model to achieve
to preserve the model with reference. Experience has shown
that a neural network model is limited by memory, which the various functions. Many successful examples show that ANN
total storage capacity is equal to the total weight of the total subsystems are similar to models built in other ways. Besides,
number of input values, and then multiplied by 8 namely: ANN subsystems outstand clear advantages in the elimination
of noise, nonlinear optimization, and exploration of multi-
The total amount of storage (bytes) = (the total weight + the variable complexity problems
total input value) * 4 * 2.
ANN MS system draws advantages on many other
As mentioned above, the neural network based on bionics excellent neural network software development platform, so
will be used as a subsystem or model. With ANN MS that it has the ability of both BP / RBF network theory-based
platform, you can certainly perform such a component-level research and of model-based application research. System
application development. For example, suppose you have
functions are mainly dependent on toolbox functions which
created a neural network, Echo, and trained it to identify 26
are associated with the BP network, and with the RBF network
English letter sounds; then have created a neural network, See,
and trained it recognize 26 English letters. If you need to in MATLAB neural network toolbox. ANN MS has
develop a video-audio sub-system, Echo See, making the connection with the Excel software from the perspective of
reading 26 letters displays literal text on the screen, just simply subsystem, and turning the developed subsystem into
combined the Echo model and the See model organically. components, as well as turning developed subsystem into
Java, VC++, VB and other high-level languages, in order to
facilitate data collection and import, component-level
IV. DISCUSSION ON APPLICATION RESEARCH
development for neural network systems, and neural network
A. Research on BP and RBF application system development.

Artificial neural network can partially simulate the human

brain work. For the future computers, and even the
B. Related Basic Application Research diagnose. Practice shows that it is feasible for the electronic
For teachers’ guidance, feed-forward neural networks, data circuit block of fault detection and diagnoses.
mining can be carried out to classification, modeling, and
relatively forecasting according to the network model and V. CONCLUSION
algorithm ANN MS system provides only the model-based As mentioned earlier, ANN artificial neural network can be
application research windows: used to solve the artificial intelligence problem without having
The first one is modeling classification window. Being used to create a true model of biological system. Although there are
the imitation of biological neural network model, here many successful examples, many problems followed,
introduce ‘modeling classification' concept, that is, the output especially on understanding and explaining of the results, have
layer nodes represent characteristics of a data group. For exposed the shortages of artificial neural networks. Artificial
example, in a course, score is divided: 100-90 points as A- neural networks built from the view of bionics package
class, 89-80 points as the B-class, 79-70 points as the C-class, themselves in the black box for the simulation of the brain
69-60 as the D-class, 60 points or less as the E-class. neurons, process and store multi-variable inputs in parallel, and
Therefore, there is one neuron in the modeling classification form models under the teacher’s signal or the error allowed to
model input layer receiving scores, and there are five neurons play some role.
in the output layer representing the A, B, C, D, E classes. As functional models, or components, ANN can be used to
A RBF network net is built with newrb function to achieve re-development for application systems so as to promote the
the goal of academic achievement classification. As the application research on ANN network, meanwhile, showing the
number of RBF neurons are increased three by three gradually, limitation of theory and application. In the summary of
the precision of 0.1 reached when the number of hidden layer previous work on neural networks, ANN MS system
neurons is 9. Then the error is 3.84671e-024, shown in Fig.3. demonstrates a research and application management system
based on BP / RBF neural networks. It can do either for
The results of fig.3 also show that the detection value of P1 theoretical research or for model-based applied research and
(0.97) through the network value of net output response is development. ANN MS has real and profound significance in
1.0760 (approximately 1) belonging to A-class, which meets the current neural network. Its results are quite considerable in
realistic expectations of the standard. the first project period, and its foreground is very bright in the
The second window is model mining Data mining rolling construction of projects.
generally passes undergoing steps of the modeling (algorithm
design), mining, evaluation and improvement; while modeled ACKNOWLEDGMENT
mining is mining after modeling, using models to find or to set Thanks to teacher Han, who taught me Neural Network
up a new relationship, using model to assist decision-making. Course, for giving me deep impression of this magic area.
The so-called data mining is a process that uses computer Thanks to my mother for her tolerant of my work.
learning technology, to automatically analyze data in the
database and extract knowledge For mimicking biological
neural network model, the learning process is the mining, REFERENCES
evaluating improving process. Modeling is a sort of network [1] http://en.wikipedia.org/wiki/
state after learning meets expectations or the allowed error. [2] Gao Jun. Principle of Artificial Neural Network and Simulation
Therefore, the window is required with strong model Instance. Beijing: Mechanical Industry Press ,2007-02:pp.43.
management capabilities, and can be marked on the models in [3] Hou Yuan-bin, Du Jing-yi, Wang Mei. Neural Network. Xi'an: Xian
Electronic Science and Technology University Press ,2007-08:pp.131.
group in order to support decision-making quickly and
effectively so as to achieve the purpose of model mining. [4] Shen Xiao-yong, Yang Bo. Database and Application. Beijing:
Tsinghua University Press ,2007-11.
The final window is fault detection and diagnosing window. [5] Han Li-qun. Artificial Neural Network Theory, Design and
The principle of the window is modeling faults: the input data Applications (2 Edition). Beijing: Chemical Industry Press ,2007-07.
is faults, while the output data is reasons, so as to detect and [6] Xiong Hao, Yan Hai-hua, Huang Yong-gang, Guo Tao, Li Zhou-jun.
Code Similarity Detection Approach Based on Back-propagation Neural
Network. Computer Science[J], Vol37, No.3:pp.159.
[7] Wang Ke-jun, Wang Ke-cheng. Neural Network Modeling, Prediction
and Control. Harbin: Harbin Engineering University Press ,1996-12:pp.4.
[8] Wang Jie, Zeng Yu, Zhang Jian-lin. MPI Runtime Parameters Tuning
Based on Neural Network on Multi-core Clusters. Computer Science[J],
Vol37, No.3:pp.229.
[9] Chen Jing-min. Data Warehouse and Data Mining Techniques(2nd
Edition). Beijing: Electronic Industry Press ,2007-11:pp.196 .
[10] Gao Ningning. The Data Modeling Mining Based on Excel. Central
University for Nationalities (Natural Science)[J], v.17 (1) :pp.49.
[11] (United States) Richard J. Roiger, Michael W. Geatz. Wengjing Nong
(translation). Data Mining Course. Beijing: Tsinghua University Press
,2003-11:pp.4.
Figure 3 Creation net process and Simulation results