|
Du bist hier: Referate Datenbank | Englisch
| Artificial Intelligence
Artificial Intelligence
S.H.Nov
1994
Artificial
Intelligence:
Today there are two main groups in the ai
research.
The ones are the logicists and the
others are the connectionists. The logicists are the traditional ai
scientists who try to create ai on computers based at the "von Neumann
model".
The connectionists try to build ai by
watching and recreating the human brain.
The
logicists:
The
History: One of the first people who
thought about
artificial intelligence was Gottfried
Leibnitz a 17`th century mathematican, who fantasised about logical
techniques.
At this time most scientists tried to find an
ultimate reasoning mechanism.
In the 50`s the most scientists recognised
that there couldn`t be a general theory of intelligence.
In the late 60`s Joseph Weitzenbaum created a
programm called "eliza" . it was the first automatet
psychiatrist.
in the mid 70`s the stanford resarch center
build a roboter called "shakey" with it`s ai programm it was barely able to
navigate an empty corridor. But it was a flop.
S.H.Nov
1994
Definition and
structure:
The most common ai programms are "expert
systems" ,this programms are desingned to capture human expertise and try to use
it to find the best logical solution.
In expert systems the programmers try to
codify
the principls of valid reasoning in form of
mathematical equations, so that the ai program sees the life as a book of rules
it has to obey.
Most expert systems use a piece of software
called "inference engine". This small program
is able to apply the rules programed in the
system to the information that was fed into the system. The software works as
long as it has rules, when there arn`t any rules left to apply then the systen
has to make it`s decision.
An example: if the season is winter and a
tree is green the system comes to the conclusion that the tree has to be
evergreen, and if the tree has a specific shape the system knows what sort of
tree it might be.
But what if the tree grows in tropical
climate and the system has no rule for such a circumstance.
And that is one of the biggest problems the
logic orientated ai has. An expert system has to be right the first time it
makes a decision because it`s unable to retract it.
The system is also unable to solve a problem
that is not in it`s "big book of rules for sucsessful and satisfied
living".
That is one of the reasons why many
scientists today believe that the final solution lies somewhere between the
logical and the connectional ai systems.
S.H.Nov
1994
Today`s usage and
future:
Today ai can be used for nearly everything.
For example the ai programm that was used for the roboter shakey (mentiond
above) has been the grandfather for some organization and planning programms
that where used in the gulf war.
These programms helped the United States to
coordinate and organize their troop movements.
American Airlines also uses such an Ai
program to help to find the best flying strategy so that they can keep up with
rival airlines.
Matsushita also uses Ai programms in their
new camcorders to cut out the jitter that is caused by shakey hands. The
programm eliminates every unexpectable or to quick motion so that only clean and
smooth motions stay on the film.
Another big company that uses expert systems
is American Express.Because about 60% of the transactations American Express has
to handle with are so common that an Ai program can do them, and the rest of 40%
are done by the ordinary personal.
In the future the Ai has great chances to
advance to the most common computer programs that we will be using in our every
day life. But we have to learn a lot if we want to use them correct and
efficient.
The probably best would be if we clould learn
to combine our human sense and the machine`s memory and logical skills. But this
is harder to do than to say because it`s already hard for us to share power with
other humans and how would some of us react if a computer program tells them
what they should do or not do. It` a hard process , but we will have to learn to
make decisions in harmony with the computers.
The
connectionists:
The
beginning: In 1908 an Italian scientist
named Cajal found
out that the neurons where the basic pieces
of our brain. Then in 1914 Adrian saw that these neutrons and the other nerves
are using litle electric impulses (about 40mV) to communicate with each other.
Many decades later in 1964 the american scientist Eccles wrote a book about
Synapses an the electrical circuits. A synapse is like a small switch that uses
chemical substances to give an electric impulse from on nerve to another. With
the help of the chemical substances our body can controll how strong such an
impulse transfered to the other nerve.
In 1943 Mc Culloch & Pitts two american
scientists wrote a work about neuronal nets and how they could be created
artificial.These two where the first who thougt about such nets and with their
book the connectionistic science has started. 1958 Rosenblatt (USA) built his
legendary "Perceptron" it was the first neuronal net, but more about that later.
After Rosenblatt many scientists tried to analyse his work. The most important
where Lettvin (1959) and Hubel & Wiesel (1962). Their book about receptive
fields started the mathematical investigations of the neuronal nets. Eleven
jears after Rosenblatt the two scientists Minsky & Pappert from the
Massecusetts Institute for Technology (MIT) had nearly killed the
connectionistic science because they proved that a retina perceptron could not
recognize every patter. (In their time they where right, but today we know that
a peceptron could recognise every patter if it only has enough
units).
But some european scientists didn`t give up,
the two most important are Tuevo Kohonen (Finnland) and Eduardo Cainello
(Italia).
Neurons and Units:
The smalest parts of neuronal nets today are
units. Units work very equal to the neurons in the human brain.
Originaldokument enthält an dieser Stelle eine Grafik!
Original document contains a graphic at this position!
As seen above a human neuron looks like
cuttlefish. It has fixed connections to other neuron`s the dentrides and it has
influenceable connections the nerves with their synapses.
The dentrides are like wires between between
two CPU`s so that the electric Impulses can`t be influenced on their way to the
neuron.
In the synapses the electric Impulses are
converted into chemical substances, when they reach the präsynaptic
membrane.
These substances have to go through the
synaptic gap.
When they reach the postsynaptic membrane
they are being reconverted into electric signals.
Now our body is able to control how much of
the substances an electic impulse of a defind strengh can create. So the body
can modify the weights between the neurons and that`s the way we are able to
learn.
The now following basic parts of modern
neuronal nets the units are quite equal to the neurons.
Originaldokument enthält an dieser Stelle eine Grafik!
Original document contains a graphic at this position!
A unit has three main parts
:
1.) the input here all input signals
are summed up and converted to the netto input.
2.) then it has an activating
function.
the activatig function says if the netto
input is high enough to create an output signal. the three most common
activating functions are shown above (Pic. 2 to 4). Picture 2 shows a sigmoide
activating function.
Picture 3 a linear and picture 4 a jump
function.
3.) the output function creates an
output signal if the activating function tells it to to so.
like the brains neurons the units are
connected with each other. They are called weights, these weights can be positiv
or negative. Positive means that they activate the next unit. Negative means
that they inhibite the next unit.
how can they
learn?:
In every neuronal net the knowlege is saved
in the weights between the units, so if you want a neuronal net to learn you
have to change the weights and if you want to save a lot of knowlege you need at
least ten million units (the human brain has about 4 billion
neurons).
If you want to change the weighs you have to
use learning rules.
Some of the most common rules: 1.) The
Hebbrule says if the units a and b where aktivated repeated very strong then you
should increase the weigths between them.
:
Originaldokument enthält an dieser Stelle eine Grafik!
Original document contains a graphic at this position!
2.) The Deltarule renders the
difference between the wanted output and the real output and creates so the
amount how much the weight should be changed.
:
Originaldokument enthält an dieser Stelle eine Grafik!
Original document contains a graphic at this position!
3.) The Backpropagation is the
Deltarule for nets with hidden layers.:
Originaldokument enthält an dieser Stelle eine Grafik!
Original document contains a graphic at this position!
it renders the faults back from one layer to
the former one.
Originaldokument enthält an dieser Stelle eine Grafik!
Original document contains a graphic at this position!
4.) competitve learning: The units in
one layer have negativ weigts to the others in that layer so that only the
strongest suvives an remains active:
Originaldokument enthält an dieser Stelle eine Grafik!
Original document contains a graphic at this position!
Some examples:
today there are quite a lot neuronal nets in
use. Sejnowski and Rosenberg build “Nettalk” which is able to
recognise and pronounce about 1000 english words with a probability thats
it’s right of 93%. It’s also able to pronounce 78% of any english
word it has never seen before. It consists of seven inputlayers each with 29
inputunits, then it has one hidden layer with 80 units and one outputlayer with
26 units. It identyfies a word by checking 7 letters. Each of the 26 outputunits
stands for one english phoneme. The net has been trained with the backpropagtion
learning rule.
The second one is the Perceptron. It has a
few sensorunits(=input), 4 or 5 associationunits and 3 or 4
responseunits(=output). In the perceptron the user is only able to change the
weights between the associationunits and the responsunits. The problem the
percepron had, was that it had not enough units, so that it wasn’t able to
recognise every inputpattern.
Future and
problems:
Today the realization of neuronal nets is a
big problem because it’s very hard to connect more than 100 units. If you
try to simulate them on normal computers you have to write a programming speech
first. It’s also very hard to programm the learning rules in the form you
want to use them.
Maybe in the future when new laser based
processors are on the marked we could be able to use the big advanteges of the
neuronal nets. With them it’s very easy to make parallel distributed
processes and content adressed memory acess.
|
PDF-Format (43,48 KB)
ZIP
Archiv (41,06 KB)
(c) 2003-2012 ontaps GmbH