Type of research: applied Duration from: 01/01/91. to 12/31/96. Papers on project (total): 32
Papers on project quoted in Current Contents: 3
Institution name: Fakultet strojarsva i brodogradnje, Zagreb (120) Department/Institute: Automation and Measurement Department Address: Ivana Lučića 5. P.O.B. 194, HR - 1000 Zagreb, CROATIA City: 10000 - Zagreb, Croatia
Communication
Phone: 385 (0) 1 6111-944
Fax: 385 (0) 1-514-535
E-mail: Vojislav.Kecman@x400.srce.hr
Summary: In recent years the importance of better
understanding, modeling and control has increased greatly in all fields of
human endevours e.g. in biology, physiology, medicine, economics, ecology
and certainly in the field of engineering, or more precisely in control,
automatization and flexible manufacturing. In technical applications the
creation of models (theoretically or by experiments) and control of the
processes and systems is an essential part of man's intelectual activity.
Throughout the past years we have been witnessing a fast development and
recognition of these fields which is related to requirements from practice
stimulated by the need for: - satisfying of quality, productivity and
enviromental demands, - responding to global undustrial competition
-responding to global industrial competition, - more information and better
understanding of processe and systems. My personal contributions, as
well as the very contributions of my younger coworkers Mr. J. Petric, Mr.
D. Majetic, Mr. M. Siroki (published in many papers, monographs and books)
are particularly strong in the field of mathematical modeling of system
dynamics, in original approach in treating the dynamics of different
processes in a unified and consistent way and in theory and application of
modern computing paradigms such as Artificial Neural Networks (ANN) and
Fuzzy Logic System (FLS). The newest field of our research (and main
part of this proposal) is connected with recent advances in ANN and FLS.
This originated from problems when it is not possible (or at least it is
very hard) to adequately represent system characteristics such as
nonlinearity, time delays, saturation or time-varying paremeters. For such,
very common, situations neural networks can be of great interest to the
dynamics and control engineers because they have potential to treat many
problems that cannot be handled by traditional analytic approaches. I am
sure that ANN and FLS with their massive parallelism, approximation,
generalization and learning capabilities can provide better solutions to
(at least some) old and new control problems. This capability of
approximation, generalization, learning and embeding of the existing human
knowledge stands behind the word intelligent in the title of this proposal.
As it is well known to all working in the computer time consumption
field of neural networks, a lot of simulation time has been spent in my
laboratory too , trying to train hundreds of network configurations for the
sake of dynamics and control of real physical systems. It has been proved
that out of many types of neural networks particularly the multilayerd
feedforward ANN with error back-propagation algorithm can be used for
simulation of dynamical systems. It is interesting to note that for
back-propagation network common sigmoidal transfer function gives much
worse results than sinusoidal transfer function. The last one has also
better properties than tangens-hyperboloid transfer function but this is
not so firm claim at this moment. Thus, for detailed experimental training
of three-therm controller (PID controller) the sinusoidal transfer function
was used. The series of simulation runs of different systems proved that
the choice of training procedure and ANN structure is of great importance
and right now a series of simulation experiments is trying to establish the
connections between the number of hidden layers and processing elements in
them with the type of dynamical systems. The world's new research is now
taking place in the field of aplication of RBF Networks as well as in
connecting RBF with FLS. Finally, our newest and the most intriguing
investigations are now in the field of design of inverse-dynamics
controller using the neural network approach. This concept of an
inverse-dynamics controller is an old one but the realization of such
controller is not an easy task, particularly when the system dynamics is
unknown and with varying and uncertain parameters (gains, poles, zeros).
Also, very strong research will take place in the field of solution of some
vision problems as well as in control of flexible robots. We plan to do
research also in the field of more deeper understanding of foundations of
ANN and FLS paradigms.
Keywords: Identification and Control of Nonlinear Systems, Artificial Neural Networks, Fuzzy Logic Systems
Research goals: Final goal of proposed research is to merge
Artificial Neural Networks and Fuzzy Logic Systems computing paradigms.
Recently, we proved a very strong result about equivalence of RBF Network
and FLS. This is a world new and powerful result and because of its
importance we wil try to track this line of research, too. The goals will
be also the development of theories, methodologies, algorithms and software
for these computing paradigms.
COOPERATION - PROJECTS
Name of project
: Neural Networks in Control Systems Name of institution: Institut fur Automatisierungstechnik,
Universitat Bremen City: Bremen, Njemačka
Name of project
: 2-08-171 ALGORITMI VOĐENJA ROBOTA I
FLEKSIBILNIH PROIZVODNIH SISTEMA Name of institution: Fakultet strojarstva i brodogradnje City: 10000 - Zagreb, Croatia
Name of project
: 2-07-176 FLEKSIBILNA AUTOMATIZACIJA PROIZVODNJE
I INDUSTRIJSKI ROBOTI Name of institution: Fakultet strojarstva i brodogradnje City: 10000 - Zagreb, Croatia Other information about the project.