- Type of paper
: Paper in journal
Title: Modeling of Dynamics and Control Systems design of
Two-Phase Circulation System of Steam Generator
- Authors:
- Kecman, Vojislav (20623)
- Petrić, Joško
- Majetić, Dubravko (162406)
- Široki, Mladen (162364)
Journal: Strojarstvo
Number: 2
ISSN: 0562-1887
Volume: 33
Year: 1991
Pages: from 117 to 124
Number of references: 16
Language: hrvatski
Summary: The paper cinsiders the development of teh mathematical
model ofsteam generator circulation loop dynamics, the linearization
oforiginal model and design of modern multivariable control system.The
limitations and benefits of linear control theory arediscused. Based on the
results of simulations of this linearmodel it is posibble to evaluate the
importance of measurement ofparticular state variables and to test the
integrity of controlsystem. Particular attention is given to the
implementation oflinear optimal controller considering the real
nonlinearenvironment consisting of bounded controlling elements.
Theinfluence of the set-points changes, feedback control matrixgains and
manipulating elements to the response of nonlinearsystem is analysed.
Keywords: Mathematical modeling, Nonlinear dynamics, Optimal controller, Steam generator
- Type of paper
: Paper in proceedings
Title: Synthesis of PID Controller by Neural Network
- Authors:
- Majetić, Dubravko (162406)
- Kecman, Vojislav (20623)
- Editors
- Kecman, Vojislav (20623)
Proceedings title: Zbornik Radova KoREMA'36
Language: hrvatski
Place: Zagreb
Year: 1991
ISBN/ISSN: 86-81571-09-5
Pages: from 55 to 58
Meeting: Sedmi simpozij o sistemima automatskog upravljanja
Held: from 04/18/91 to 04/20/91
Summary: A paper presents synthesis of PID controller
byHetero-associative, four-layer, back-propagation neural network.One of
the possible application of neural networks in the fieldof automatic
control is discussed.
Keywords: Neural network, Error back-propagation, PID controller
- Type of paper
: Paper in proceedings
Title: Neural Network as PID Controller
- Authors:
- Majetić, Dubravko (162406)
- Editors
- Srb, Neven
Proceedings title: Zbornik radova 3.SONT
Language: hrvatski
Place: Zagreb
Year: 1991
ISBN/ISSN: 86-81997-02-5
Pages: from 62 to 65
Meeting: Četvrti međunarodni simpozij o novim tehnologijama
Held: from 10/15/91 to 10/17/91
Summary: The learning procedure of Hetero-associative,
four-layer,back-propagation neural network for learning the PID algorithm
ispresented. One of the possible applications of neural network inthe field
of automatic control is discussed.
Keywords: Neural network, PID Controller, Learning procedure, Recall procedure
- Type of paper
: Paper in proceedings
Title: Neural Network as Ideal Controller In Open-Loop Control
- Authors:
- Majetić, Dubravko (162406)
- Vinković, Mladen
- Editors
- Novaković, Branko (33541)
Proceedings title: Zbornik radova KoREMA'38
Language: hrvatski
Place: Zagreb
Year: 1993
ISBN/ISSN: 953-6037-00-9
Pages: from 450 to 454
Meeting: 38. Međunarodni godišnji skup KoREMA : Sekcija Automatsko upravljanje
Held: from 04/26/93 to 04/28/93
Summary: As neural networks shows very good results during the
learningand recall process od dinamics behaviours with distinctderivation
properties, which are the main characteristic of idealcontrollers, the
possible application of neural network in theconcept of ideal controller is
discussed.
Keywords: Neural network, Inverse-dynamics Controller, Open-Loop Controll
- Type of paper
: Paper in proceedings
Title: Simulation of Dinamic Elements by Neural Network
- Authors:
- Majetić, Dubravko (162406)
- Editors
- Kecman, Vojislav (20623)
Proceedings title: Zbornik radova KoREMA'37
Language: hrvatski
Place: Zagreb
Year: 1992
ISBN/ISSN: 86-81571-11-7
Pages: from 46 to 49
Meeting: 37. Međunarodni godišnji skup KoREMA - 8. Simpozij o sistemima automatskog upravljanja
Held: from 04/26/92 to 04/29/92
Summary: In this paper an approach to the learning procedure of
errorback-propagation neural network for learning some basic
dynamicbehaviours such as P1, PD, I0 and P2 is presented. The number
ofprocessing elements in input and output layer is defined bydicrete form
of dynamic element. Consideration is given tocertain problems wich are
connected with discrete form of dynamicelement and which occurs in recall
procedure of trained neuralnetwork.
Keywords: Neural network, Error back-propagation, dicrete form, learning procedure, recall procedure
- Type of paper
: Paper in proceedings
Title: Neural Network Image Processing Based on Textural Features
- Authors:
- Široki, Mladen (162364)
- Editors
- Kecman, Vojislav (20623)
Proceedings title: Zbornik radova KoREMA'37
Language: hrvatski
Place: Zagreb
Year: 1992
ISBN/ISSN: 86-81571-11-7
Pages: from 60 to 64
Meeting: 37. Međunarodni godišnji skup KoREMA - 8. Simpozij o sistemima automatskog upravljanja
Held: from 04/26/92 to 04/29/92
Summary: Neural network was used in recognition three different
groups ofobjects with similar geometrical shape, but different texture.They
included mushrooms, biscuits and pieces of rounded paper.Recognition is
based on textural features extracted fromco-occurrence matrices. The
overall identification accuracy wasmore than 99%.
Keywords: Neural network, Back-Error propagation, Co-occurrence matrices, Textural features
- Type of paper
: Paper in proceedings
Title: On the Realation Between the Cost Function and the
Output-Layer Neurons Activation Function
- Authors:
- Kecman, Vojislav (20623)
Proceedings title: Procceedings of '15. Kolloquium der Automatisierungstechnik
Language: engleski
Place: Bremen, Njemačka
Year: 1993
Pages: from 0 to 0
Meeting: 15' Koloquium der Automatisierunhstechnik
Summary: This paper examines the relation between the cost function
andthe activation function (AF) of the output-layer neurons ofneural
networks. It is important to realize that the well formedcost (or error)
function, when training artificial neuralnetworks (ANN) using Error Back
Propagation algorithm, cangreatly reduce the learning time. This cost
function isdetermined by final task the ANN is applied for i.e., by
probabilistic model for how the training data are being generated. At the
same time, the type of AF of the output-layer neurons is not independent
but rather connected with the type ofthe very cost function used.
Keywords: Cost (error) function, Activation function, Neural Networks
- Type of paper
: Paper in proceedings
Title: EBP Work With Hard Limiters
- Authors:
- Kecman, Vojislav (20623)
Proceedings title: Proceedings of International Conference on Artificial Neural Networks
Language: engleski
Place: Amsterdam, Nizozemska
Year: 1993
Pages: from 0 to 0
Meeting: International Conference on Artificial Neural Networks
- Type of paper
: Paper in proceedings
Title: Dynamic Properties and Control of the Flexible
Manipulators
- Authors:
- Petrić, Joško
- Kecman, Vojislav (20623)
- Editors
- Cebalo, Roko
Proceedings title: Zbornik radova Suvremeni trendovi proizvodnog strojarstva
Language: hrvatski
Place: Zagreb
Year: 1992
Pages: from 143 to 151
Meeting: Suvremeni trendovi proizvodnog strojarstva
Held: from 05/09/92 to 05/09/92
Summary: Flexible manipulators become one of the most interesting
field ofrobotics today. In this paper some basic dificulties, phenomenaand
issues connected by the approach to the manipulators aselastic bodies are
considered. Mention is also made about thequestion - is it necessary to
control the motion of themanipulators including flexibility of their links
into thefeedback?!
Keywords: Flexible Manipulator, Feedback Control, State-Space Control
- Type of paper
: Paper in proceedings
Title: Nonminimum-Phase Property of the Elastic Mechanical
Structures
- Authors:
- Petrić, Joško
- Editors
- Crnošija, Petar
Proceedings title: Zbornik radova KoREMA'93
Language: hrvatski
Place: Zagreb
Year: 1993
ISBN/ISSN: 953-6037-00-9
Pages: from 715 to 719
Meeting: 38. Međunarodni godišnji skup KoREMA : Sekcija Automatsko upravljanje
Held: from 04/26/93 to 04/28/93
Summary: In this paper nonminimum-phase property of the
noncolocatedcontrolled elastic mechanical structures is demonstrated on
anexample of a single-link flexible manipulator. This property,absent in
the classical control approach of rigid manipulators,is cause of the
detoriated capability of path following andprecise positioning of the
flexible manipulators. Therefore,herein the attention will be given on the
characteristics of themanipulators depending on the actuator and sensor
location alongthe link, and some connected control problems will be
alsoconsidered.
Keywords: Nonminimum-Phase Property, Elastic Body, Flexible Manipulator
- Type of paper
: Paper in proceedings
Title: On the Control of the Flexible Manipulator Arm
- Authors:
- Petrić, Joško
- Kecman, Vojislav (20623)
Proceedings title: IFAC Symposium on Robot Control - SYROCO'91
Language: engleski
Place: Wien, Austrija
Year: 1991
Pages: from 325 to 330
Meeting: IFAC Symposium on Robot Control
Held: from 09/16/91 to 09/18/91
Summary: The paper considers the control of a class of robots whose
armflexibility, due to light construction and the demand for fastrobot
motion, is not negligible. The mathematical model ofdynamics of one
translational joint and one flexible armmanipulator with the payload on its
end is developed. On thebasis of the mathematical model developed, an
optimal linearquadratic state regulator is designed. In the paper the
influenceof additional control action, using the states of vibration
modesof the manipulator arm, is analysed.
Keywords: Robot, Flexible Manipulator, Modeling, State Space, Optimal Control
- Type of paper
: Paper in proceedings
Title: State-Space Based Control System Design for Flexible
Manipulators
- Authors:
- Kecman, Vojislav (20623)
- Petrić, Joško
- Petrić, Joško
- Editors
- Cebalo, Roko
Proceedings title: Fachtagugn Automatisierung Dresden'92
Language: engleski
Place: Dresden, Njemačka
Year: 1992
Pages: from 146 to 152
Meeting: Fachtagung Automatisierung
Held: from 02/20/92 to 02/21/92
Summary: The paper considers the joint-states feedback controland
the control which also involves feedback of vibration statesin the flexible
single link manipulator case. An optimal linearquadratic state controller
is applied and the advances in thecase of extended, noncolocated feedback
control are discussed.
Keywords: Flexible Manipulator, State-Space, Optimal LQ Controller
- Type of paper
: Paper in proceedings
Title: Applicatin of Artificial Neural Networks in Robotics
- Authors:
- Novaković, Branko (33541)
- Editors
- Cebalo, Roko
Proceedings title: Zbornik radova CIM'93
Language: hrvatski
Place: Zagreb
Year: 1993
Pages: from 1 to 8
Meeting: 2. Međunarodno savjetovanje proizvodnog strojarstva
Held: from 11/18/93 to 11/18/93
Summary: Starting with the general artificial neural network
model,described as discrete time nonlinear system, the feedforwardneural
network for robot control has been derived. This networkis trained to
imitate an adaptive nonlinear robot controlalgorithm based on the dynamics
of the full robot model of RRTR-structure. As the result we have the neural
network , which cancompute both the nominal and feedback robot control,
using theparallel processing.
Keywords: Artificial neural networks, idustrial robot control, adaptive control, fast learning algorithms.
- Type of paper
: Paper in proceedings
Title: Neural Networks and Control in Robotics and FMS :
Conceptual overview
- Authors:
- Novaković, Branko (33541)
- Lišćić, Igor
- Furač, Zvonimir
- Editors
- Srb, Neven
Proceedings title: Zbornik radova 4.SONT
Language: hrvatski
Place: Zagreb
Year: 1993
ISBN/ISSN: 86-81997
Pages: from 31 to 34
Meeting: Četvrti međunarodni simpozij o novim tehnologijama
Held: from 10/25/93 to 10/27/93
Summary: A conceptual overview of the state of the art in theory
andapplications of neural networks in automatic control and roboticsis
given in this paper. Besides, some new methods are proposed.The control
methods of flexible manufacturing cell are discussedat the end of this
paper.
Keywords: Neural networks, industrial robots, flexible manufacturing systems, a conceptual overview.
- Type of paper
: Paper in proceedings
Title: An Application of Neural Networks in the Control of the
Industrial Robots
- Authors:
- Lišćić, Igor
- Novaković, Branko (33541)
- Furač, Zvonimir
- Editors
- Novaković, Branko (33541)
Proceedings title: Zbornik radova KoREMA'38
Language: hrvatski
Place: Zagreb
Year: 1993
ISBN/ISSN: 953-6037-00-9
Pages: from 458 to 461
Meeting: 38. Međunarodni godišnji skup KoREMA : Sekcija Automatsko upravljanje
Held: from 04/26/93 to 04/28/93
Summary: This paper presents the realisation of one link control
ofindustrial robot, using artificial neural networks (ANN). Thecontrol
system is created by two ANN, where the first one computsnominal control,
and the second network is a nonlinear controllerof position and velocity of
robot.
Keywords: Artificial neural networks, nominal control, controller of position and velocity of robot.
- Type of paper
: Paper in proceedings
Title: Aquisition Software Possibilities for Vibration
Measurement
- Authors:
- Grilec, Josip (14562)
- Editors
- Šurina, Tugomir
Proceedings title: Zbornik radova BIAM'92
Language: hrvatski
Place: Zagreb
Year: 1993
ISBN/ISSN: 86-81571-12-5
Pages: from 44 to 49
Meeting: XI Međunarodno savjetovanje BIAM'92
Held: from 06/16/92 to 06/17/92
Summary: This article gives the possibilities of the programme which
notonly supports the sampling circuits, but can also quickly performa
series of activities concerning processing, graphicalpresentation and
storing of vibration signals on a personalcomputer.
Keywords: Acquisition software, vibration measurement, signal sampling.
- Type of paper
: Paper in proceedings
Title: Criterion Systematisation for Sensor Selection in Flexible
Manufacturing System
- Authors:
- Grilec, Josip (14562)
- Editors
- Srb, Neven
Proceedings title: Zbornik radova 3.SONT
Language: hrvatski
Place: Zagreb
Year: 1991
ISBN/ISSN: 86-81997-02-5
Pages: from 66 to 68
Meeting: Treći međunarodni simpozij o novim tehnologijama
Held: from 10/15/91 to 10/17/91
Summary: This paper discusses procedures for choosing sensors of
flexiblemanufacturing systems and a systematisation of
characteristicfeatures with restrictions. As one has to be aware of
somepossible environmental influences, a number of desirablecharacteristic
features according to different criterions arepresented. Because of the
poor standards special stress is put onthe importance of the right choice
of sensors.
Keywords: Criterion systematisation, environmental influences, flxible manufacturing systems, sensor selection.
- Type of paper
: Paper in proceedings
Title: Application of Parallel Processing on Robot Dynamic
Control
- Authors:
- Zorc, Davor (111286)
- Editors
- Srb, Neven
Proceedings title: Zbornik radova 3.SONT
Language: hrvatski
Place: Opatija
Year: 1991
ISBN/ISSN: 86-81997-02-5
Pages: from 83 to 87
Meeting: Treći međunarodni simpozij o novim tehnologijama
Held: from 10/15/91 to 10/17/91
Keywords: Parallel Processing, Process Alocation, Robot Control
- Type of paper
: Paper in proceedings
Title: Exploiting The Structural Equivalence of Leraning Fuzzy
Systems and Radial Basis Function Neural Networks
- Authors:
- Kecman, Vojislav (20623)
- Pfeifer, Bernard-Markus
Proceedings title: Proceedings of Sec.Eur. Congress on Intell. Techniq. and Soft Comp., EUFIT
Language: engleski
Place: Aachen, Njemačka
Year: 1994
Pages: from 58 to 66
Meeting: Seccond Europian Congress on Intell. Techniq. and Soft Comp., EUFIT'94
Summary: This paper shows when and how the Learning of Fuzzy Rules
(LFR) from numerical data is equal to the training of Radial Basis Function
(RBF) or regularization networks. Although both approaches have their
origin in different paradigms of "intellignet" information processing, it
is demonstrated that the mathematical structure is the same and they share
the property of being "universal approximators" of any real coninuous
function on a compact set to arbitrary accuracy. In the LRF algorithm
proposed here the subject of learning are the rule conclusions, i.e. the
membership functions of output attributes in the form of singletons. For
fixed number, location and shape of input membership functions LFR and RBF
training are least squares optimization problems linear in the unknown
parameters, and in this case the solution boils down to the pseudo
inversion of one rectangular matrix. As an application example the dynamic
identification of a nonlinear real world system from noisy measurement data
will be performed with LRF and with "Soft" RBF.
Keywords: neuro fuzzy systems, fuzzy identification, RBF networks
- Type of paper
: Paper in proceedings
Title: Dynamic Neurla Network for Adaptive Nonlinear Robot
Control
- Authors:
- Majetić, Dubravko (162406)
Proceedings title: Zbornik radova 40. jubilarni godišnji skup KoREMA
Language: hrvatski
Place: Zagreb
Year: 1995
ISBN/ISSN: 953-6037-08-4
Pages: from 498 to 501
Meeting: 40. jubilarni godišnji skup KoREMA
Held: from 04/19/95 to 04/21/95
Summary: The paper presents a possibility of application of a
multilayer feedforward dynamic neural network in robot control. The hidden
layer consists of dynamic neurons. Eash dynamic neuron posseses one dynamic
element with two poles and two zeros followed by hyperbolic tangent
activation function. The output layer consists of static neurons with
linear activation function. This neural network is trained to imitate an
adaptive nonlinear robot control algorithm, based on inerse dynamics of the
full robot model of RRTR-structure. For that purpose the error
backpropabation supervised learning algorithm with stochastic learnig is
developed and neurons weights and coefficients of dynamic neurons are
learned simultaneously.
Keywords: dynamic neural network, adaptive nonlinear robot control
- Type of paper
: Paper in proceedings
Title: Eksperimental Model of the Flexible Robot
- Authors:
- Petrić, Joško
- Deur, Joško
Proceedings title: Zbornik radova 12. međunarodnog savjetovanja BIAM'94
Language: hrvatski
Place: Zagreb
Year: 1994
Pages: from F15 to F18
Summary: In this paper an experimental model of the robot with one
rotation joint and a very flexible link is presented. A mathematical model
of robot dynamics is verified, and also experiments on the flexible robot
control was done on this laboratory facility. A brief research description
is given in the paper. Significance and reasons of research on the field of
flexible robots, from authors point of view, are also explained here.
Keywords: flexible robot, eksperimental model, control
- Type of paper
: Paper in proceedings
Title: State of the Art in Neural Networks Applications to Robot
Control
- Authors:
- Novaković, Branko (33541)
- Editors
- Novaković, Branko (33541)
Proceedings title: Zbornik radova 40.Jubilarni godišnji skup KoREMA
Language: engleski
Place: Zagreb
Year: 1995
ISBN/ISSN: 953-6037-08-4
Pages: from 487 to 491
Meeting: 40. jubilarni godišnji skup KoREMA
Held: from 04/19/95 to 04/21/95
Summary: In the paper an overview of the applications of
feedforward, feedback, and adaptive Neural Networks (NN's) to robot control
is presented. The following items are pointed out :(i) feedforward NN for
adaptive nonlinear robot control, (ii) a neurla network for
multi-manipulator system control, (iii) a nonnumerical environment
modelling, (iii) adaptive NN for adaptive nonlinear robot control, (iiii)
NN controllers for mobile robots, (iiiii) a global linarising control using
recurrent NN, and (iiiiii) making of self-robot control more "human-like".
- Type of paper
: M.A.
Title: Contribution to the Use Neural Networks in the Systems of
Automatic Control
- Authors:
- Kecman, Vojislav (20623)
- Majetić, Dubravko (162406)
Faculty: Strojarstva i brodogradnje Zagreb
Date of defense: 02/14/92
Language: hrvatski
Number of pages: 130
Summary:
In the M.S. thesis, entilted "THE CONTRIBUTION TO THE USE NEURALNETWORKS IN
THE SYSTEMS OF AUTOMATIC CONTROL" the learningprocedure of
hetero-associative, multilayer, backpropagationneural network for learning
some basic dynamic behaviours such asP1, P2, D1, PD and I0 is presented.
The analysis of the structure of proposed neural network is alsogiven, with
special emphasis on the number of hidden layers,number of processing
elements in output, hidden and input layer,and on the content of learning
file.
The optimal setting of learning coefficients C1 and C2 inlearning rule
during the learning procedure is proposed.
The processing element transfer function analysis has also beencarried out
and in the sense of faster decreasing of learningerror the best processing
element transfer function is proposed.
Possible applications of neural networks in the field ofautomatic control
is also discussed. Consideration is given tocertain other questions which
may be resolved by further researchand which, at the same time, set the
guidelines for continuedresearch in this area.
Keywords: Neural networks, Number of hidden layers, Number of Neurons, Activation Function, Learning Coefficients
- Type of paper
: M.A.
Title: Contribution to the Development of Inteligent Vision
System
- Authors:
- Kecman, Vojislav (20623)
Faculty: Strojarstva i brodogradnje Zagreb
Date of defense: 11/19/92
Language: hrvatski
Number of pages: 143
Summary: In this report one approach of the Automatic Intelligent
VisionSystem will be presented. The Vision System described in thiswork
consists of the following stages : image capture, low levelimage
processing, features extraction and objects recognition.The special
attention in this work is given to the featuresextraction and objects
recognition by the Back-Error propagationneural network. The objects
recognition is based on the texturalfeatures extracted from Gray-Tone
Spatial Dependence matrices(Co-occurrence matrices). The most significant
textural featureswere selected by Stepwise discriminant analysis (Wilks
method).Objects were classified in their groups by Back-Error
propagationneural network.
The transputer based Vision system is designed and developed.
Thetransputers offer new possibilities in designing and
programmingcost-effective concurrent system. It is possible to build
aconcurrent system which will perform a real-time imageprocessing.
The Vision system is used in classifying three different groupsof objects
with similar geometrical shapes but different surfacetexture. They included
mushrooms, biscuits and pieces of roundedpaper. The objects were divided in
two sets - training set andtest set. The number of textural features were
extracted fromCo-occurrence matrices of objects from training set. The
mostsignificant features were selected by Stepwise discriminantanalysis.
These features were used for the Back-Error propagationneural network
training. The network were trained until the errorfor entire set was less
then 0.01. The Vision System was testedby the test set. The overall
identification accuracy produced wasmore then 99 % . These results indicate
that the neural networkclassification based on the textural features
extracted fromCo-occurrence matrices can be very promising method, and
probablycan find wide applications in the practice.
Keywords: Vision System, Neural Network, Back-Error propagation, Co-ocurrence matrices, Textural features
- Type of paper
: Mentorship
Title: Contribution to the Use Neural Networks in the Systems of
Automatic Control
Faculty: Strojarstva i brodogradnje Zagreb
Mentor: ADLER NEVENKA
Date of defense: 02/14/92
Number of pages: 130
Author: Majetić mr.sc. Dubravko
Degree level: M.A.
- Type of paper
: Mentorship
Title: Contribution to the Development of Inteligent Vision
System
Faculty: Strojarstva i brodogradnje Zagreb
Mentor: ADLER NEVENKA
Date of defense: 11/19/92
Number of pages: 143
Author: Široki mr.sc. Mladen
Degree level: M.A.
- Type of paper
: Survey/Study
Title: The Automatic Harvesting of the Mushrooms - the Grasping
Mechanism
- Authors:
- Široki, Mladen (162364)
Ordering party: The University of Wollongong, Department of Electrical and Computer Engineering
Institution depot: The University of Wollongong, Department of Electrical and Computer Engineering
Year: 1991
Number of pages: 23
Language: engleski
Summary: In this report the design, development and implementation
of amushroom grasping mechanism will be presented. This is the partof the
Automatic Harvesting of the Mushrooms project, which aimsto design and
develop Automatic Mushroom Harvesting which willbe able to singularise,
recognize, grasp and harvest themushrooms without demaging them.
Keywords: Grasping Mechanism, Mushroom harvesting, Automatic Mushroom Harvesting Machine
- Type of paper
: Survey/Study
Title: The Automatic Harvesting of the Mushrooms - the Vision
System
- Authors:
- Široki, Mladen (162364)
Ordering party: The University of Wollongong, Department of Electrical and Computer Engineering
Institution depot: The University of Wollongong, Department of Electrical and Computer Engineering
Year: 1992
Number of pages: 20
Language: engleski
Summary: In this report using of textural features extracted
fromCo-occurrence matrices will be presented. The method is intendedto be
used in an automatic Mushroom Harvesting Machine.Classification is
performed by Back-Error Propagation NeuralNetwork.
Keywords: Vision system, Textural features, Co-occurrence matrices, Neural Network, Back-error propagation
- Type of paper
: Survey/Study
Title: Application of Artificial Neural Networks for
Identification of System Dynamics
- Authors:
- Kecman, Vojislav (20623)
Ordering party: Massachusetts Institue of Tecmnology, Department of Mechanical Engineering
Institution depot: Massachusetts Institue of Tecmnology, Department of Mechanical Engineering
Year: 1993
Number of pages: 74
Language: engleski
Summary: This report discusses the application of Feed-Forward
Multy-LayerArtificial Neural Networks in the field of system
dynamicsidentification. ANNs are presented to be general approximators
ofmultivariate functions. In this fremwork, identification is doneusing
mathematical representation of system dynamics as mappingof time histories
of past (and current) input and past outputsignals of dynamic system into a
current output signal. The imactof the number of hidden layers, number of
neurons in them, typeof activation function in neurons and training
function onaccuracy and generalization capabilities of ANN model of
dinamicsystem is analyzed. Basic of Regularization (Radial BasisFunction)
Networks are introduced as well as the survey ofcurrent relevant literature
from the field is given.
Keywords: Neural networks, Identification, System Dynamics Modeling
- Type of paper
: Survey/Study
Title: Approximation Capabilities of ANNs as Function of the
Number of Hidden Layer Neurons and the Type of Activation Function
- Authors:
- Kecman, Vojislav (20623)
- Majetić, Dubravko (162406)
Ordering party: Massachusetts Institue of Tecmnology, Department of Mechanical Engineering
Institution depot: Massachusetts Institue of Tecmnology, Department of Mechanical Engineering
Year: 1993
Number of pages: 19
Language: engleski
Summary: This paper discusses the application of Feed-Forward
Multy-LayerArtificial Neural Networks in the field of system
dynamicsidentification. ANNs are presented to be general approximators
ofmultivariate functions. Identification is done using
mathematicalrepresentation of system dynamics as mapping of time histories
ofpast (and current) input and past output signals of dynamicsystem into a
current output signal. The imact of the number ofhidden layers, number of
neurons in them, type of activationfunction in neurons and training
function on accuracy andgeneralization capabilities of ANN model of dinamic
system isanalyzed.
Keywords: Approximation capabilities, Hidden layer, Activation function
