Download Tehnici de optimizare – Programare Genetica

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts

Donald O. Hebb wikipedia , lookup

Neuroeconomics wikipedia , lookup

Neural oscillation wikipedia , lookup

Neural coding wikipedia , lookup

Optogenetics wikipedia , lookup

Connectome wikipedia , lookup

Neuroanatomy wikipedia , lookup

Artificial consciousness wikipedia , lookup

Embodied cognitive science wikipedia , lookup

Synaptic gating wikipedia , lookup

Neural modeling fields wikipedia , lookup

Central pattern generator wikipedia , lookup

Neuropsychopharmacology wikipedia , lookup

Biological neuron model wikipedia , lookup

Existential risk from artificial general intelligence wikipedia , lookup

Philosophy of artificial intelligence wikipedia , lookup

Ethics of artificial intelligence wikipedia , lookup

Mind uploading wikipedia , lookup

Holonomic brain theory wikipedia , lookup

Machine learning wikipedia , lookup

History of artificial intelligence wikipedia , lookup

Development of the nervous system wikipedia , lookup

Artificial general intelligence wikipedia , lookup

Pattern recognition wikipedia , lookup

Neural engineering wikipedia , lookup

Metastability in the brain wikipedia , lookup

Catastrophic interference wikipedia , lookup

Artificial neural network wikipedia , lookup

Artificial intelligence wikipedia , lookup

Convolutional neural network wikipedia , lookup

Nervous system network models wikipedia , lookup

Recurrent neural network wikipedia , lookup

Types of artificial neural networks wikipedia , lookup

Transcript
Leila BARDAŞUC & Andrei POPESCU – Artificial Neural Networks
Artificial Neural Networks
Leila BARDAŞUC1 & Andrei POPESCU2
1- Faculty of Economics, Alma Mater University of Sibiu, 57 Somesului Street, 550003 Sibiu, Romania
Phone/Fax: +40 269 250008, E-mail: [email protected]
2 - Polytechnic University of Bucharest, Splaiul Independentei, No. 313, Sector 6, Bucharest
Phone/Fax: +40 021 402 91 79, E-mail: [email protected]
Abstract
You do not have to look far and we can see that every electronic device is built around at least one
processor which is a miniature brain, in the most simplified way (borrows a few features that help to
control that device). Scientists try to simulate a brain to fully comply with the biological model from which
it draws its inspiration. There are a lot of researchers who make great efforts in this direction hoping that
soon they will have noticeable results. The paper tries to put in the centre of our analysis the artificial
networks, with its advantages and disadvantages.
Keywords: neural networks, artificial intelligence, learning algorithms
Rezumat
Nu trebuie să căutăm mult şi ne putem da seama că fiecare dispozitiv electronic este construit în jurul a cel
puţin unui processor ce reprezintă un creier în miniatură, în modul cel mai simplificat (împrumutâd doar
câteva funcţii ce îl ajută să controleze acel dispozitiv). Se încearcă să se simuleze un creier care să respecte
întocmai modelul biologic din care se inspiră. Sunt o mulţime de cercetători care depun eforturi mari în
această direcţie sperând că în curând vor avea rezultate notabile. Articolul încearcă să pună în central său
analiza reţelei artificial neuronale, cu avantajele şi dezavantajele ei.
Cuvinte cheie: reţele neurale, inteligenţa artificială, algoritm de învăţare
philosophers, engineers, etc. have worked tirelessly
to find the components, structure, functionality on
the one hand, and on the other, they studied the
changes occurred due to the interaction with the
environment by applying stimuli of different types
(ex. sociable, chemical, etc.).
Introduction
Since ancient times, the human mind fascinated with
its complexity all generations until today. By
definition, one of the major differences between
humans and the rest of living beings on earth is that
man has reason, which makes it high rated and to be
placed on the uppermost stage of biological
evolution.
In the lines above we mentioned a profession, which
at first glance, if superficially analyzed, not related
to the study of the mind, namely the profession of
being an engineer. On the contrary and indeed we
should not get in too depth inside the problem to
discover that they have a very important role in what
This reasoning, which represents their capacity to
think, to make their own decisions, to have feelings,
etc. is due to the brain. These are some of the
foundations that were the motivation for the brightest
people to try to unravel the mystery that lies behind
the mind. Doctors, psychologists, biologists,
27
Sibiu Alma Mater University Journals – Series C. Social Sciences – Volume 6, no. 1 / 2013
is supposedly researching and trying to elucidate the
mysteries of the mind.
head neuron, dendrites, and the axon terminal
buttons. They are interconnected to a form of an
extensive network that electrochemical processes
information. All electrochemical and electrical
impulses are transmitted by neurons with synapses.
We mention only the medical equipment that they
designed and we can see the contribution that
engineers bring in to this field. What makes it even a
more special connection between this area and
engineering is that the relationship runs both ways:
the engineers come up with solutions regarding
equipment which facilitates the study of this
problem, while studying this field provides
biological models underlying the various
technologies that engineers create and are meant to
provide further responses to different problems and
eases our daily life.
Starting from this biological example, it reached a
simplified mathematical model of a neuron model of
artificial neural network in the following way:
dendrites represent the input variables, while the
axon is represented by the output variable. The body
of the neuron is symbolized by a weighted
summation, whose role is to attribute each entry with
a specific weighting factor and then realize their sum
and the activation function which is a function that
depends on the result of the adder. For the model to
be complete, we must also add a constant value
representing the amount of polarization of the neuron
that tells us if the neuron is active or not.
Analyzing the method
Surely you've heard the term artificial intelligence;
it's about the ability of a machine (computer, robot,
electronic part, etc.) to behave intelligently.
(Pătraşcu, M. 2013) To this end, Alan Turing created
a test that can classify whether a machine is
intelligent.
Starting from this model, we can create an artificial
neural network which has a train to meet the purpose
for which it was designed, for example, recognize a
set of objects, a program created to provide inputs
using genetic programming, can take some decisions
by different script writers, etc..
The test was the following: We have a person and a
machine to be tested. In another room, there is a
panel you have to ask questions to the two
participants in the test. A jury gets the answers to
each question in writing in order to be analyzed. A
machine is considered to have passed the Turing test
if the final jury cannot decide which of the two the
person is and which is the machine is. Artificial
intelligence is used in various games that made
characters to have some different behaviour
adjustable to a single variable; it is present in
different software programs, in some search motors,
etc.
The main properties of an artificial neural network
are:
1.
2.
3.
4.
Learning capacity
Functional approximation capacity
Distributed processing of information
High processing speed in real
(Dumitrache Ioan, 2010)
time
To design an artificial neural network is required to
specify the following aspects:
1. Number of neurons
2. Interconnection mode
3. The relationship between the input neuron
signals
4. How the signals passes from input to output
5. The learning method used for training the
network
Next, we want to focus on a specific type of artificial
neural networks, algorithms namely having as a
starting point the same biological model of the brain
and the nervous system.
The nervous system consists of a large number of
neurons, about 100 billion. Each of them consists of
28
Leila BARDAŞUC & Andrei POPESCU – Artificial Neural Networks
6. Activation functions
7. The number of entering layers, hidden ones
and exit ones. (Ibidem 2)
Learning algorithms
Next we focus a bit on the learning algorithms
(methods) (training) for artificial neural networks.
This concept aims to adjust the weights and the
polarization so that the input can get outputs for
which we designed this network for. It is a very
important phase in the economy of these types of
algorithms.
The basic principle that learning algorithms work is:
apply a set of artificial neural network inputs and
weights depending on polarization neurons, to obtain
an output data set. The next step requires that once
we got a set of outputs (depending on the teaching
method chosen) the network to be travelled in
reverse, and at this step to modify the polarization
and the weights so that the next scroll goes directly
to the network, for the same data source, to set out
much closer to the desired output data.
The two categories in which learning algorithms are
divided are:
1. Supervised learning
2. Unsupervised learning
The first method is represented by training artificial
neural networks in which we specify input data
associated with output data that we want to get after
processing the input. In this case the network
weights are adjusted so that the errors of the
specified inputs and outputs will fit within the
desired accuracy of the build network. In short we
say we have an observer (represented by the error
value) that tells us when our input data produces
outputs that we have imposed (depending on the
precision selected).
The second category, unsupervised learning, does
not use information about outputs, which are not
even specified, the algorithm tries to detach itself by
defining characteristics of output data. Such
networks link the input weights of a prototype
algorithm and considers them being determined
outputs. These weights are modified as they create
new prototypes which are leading to the desired
output. These algorithms use methods of selforganization inspired by the way the brain works
when it makes associations.
Conclusions
Artificial neural network architecture has several
advantages, such as:
-
flexibility
they easily adapt to changes
they have a high tolerance to noise
the final solution is not affected by errors
occurring in the process
it is a robust method that manages well the
uncertainties
has the ability of self-organization
they can create and protect their information
in the organization
shows a high degree of parallelism.
Obviously this presents a number of disadvantages
too, because they are not easily trained, require a
large number of "training" hours for the network to
be able to act for what it was designed for, cannot
detect singular elements other than those for which it
has been accustomed (3), it's hard to debug during
operation and it is not scalable.
Considering this aspects, we highlight some
applications where we can meet these artificial
neural networks: prediction of semi-random
processes, e.g. in economy, development of
processes affected by uncertainties in decision
making, risk analysis, recognition of patterns such as
facial identification, voice, etc., to identify different
objects that have certain characteristics, we can
generate the set of primitive variables obtained by
genetic programming algorithms.
References
Pătraşcu, M. 2013. Artificial Intelligence - Lecture Notes,
Department of Automatic Control and Systems
Engineering, University Politehnica of Bucharest.
29
Sibiu Alma Mater University Journals – Series C. Social Sciences – Volume 6, no. 1 / 2013
Dumitrache Ioan, 2010, Ingineria Reglarii Automate Volumul 2, Department of Automatic Control and
Systems Engineering, University Politehnica of
Bucharest.
http://pages.cs.wisc.edu/~bolo/shipyard/neural/local.html
30