Artificial Intelligence - Rational Team for Technology LTD

advertisement
Artificial Intelligence
Lecture 2:Knowledge Representation I
Faculty of Mathematical Sciences
4th
5th IT
Elmuntasir Abdallah Hag Eltom
Lecture Objectives
[Part I: Chapter 2]
1. Look at some of the arguments against strong AI (the
belief that a computer is capable of having mental
states).
2. Look at the prevalence of Artificial Intelligence today
and explain why it has become such a vital area of
study.
3. Look at the extent to which the Artificial Intelligence
community has been successful so far in achieving the
goals that were believed to be possible decades ago. In
particular, we will look at whether the computer HAL in
the science fiction film 2001: A Space Odyssey is a
possibility with today’s technologies.
Lecture Objectives
[Part II: Chapter 3]
• Discuss representations. The reason for this is that in
order for a computer to solve a problem that relates to
the real world, it first needs some way to represent the
real world internally. In dealing with that internal
representation, the computer is then able to solve
problems.
• Introduce a number of representations, such as semantic
nets, goal trees, and search trees.
• Explains why these representations provide such a
powerful way to solve a wide range of problems.
• Introduce frames and the way in which inheritance can
be used to provide a powerful representational system.
“The limits of my language mean the limits of my world.”
-Ludwig Wittgenstein.
The Chinese Room
• The American philosopher John Searle has argued
strongly against the proponents of strong AI who believe
that a computer that behaves sufficiently intelligently
could in fact be intelligent and have consciousness, or
mental states, in much the same way that a human
does.
• One example of this is that it is possible using data
structures called scripts to produce a system that can
be given a story (for example, a story about a man
having dinner in a restaurant) and then answer questions
(some of which involve a degree of subtlety) about the
story. Proponents of strong AI would claim that systems
that can extend this ability to deal with arbitrary stories
and other problems would be intelligent.
“The limits of my language mean the limits of my world.”
-Ludwig Wittgenstein.
The Chinese Room
Searle’s Chinese Room experiment was based on this idea
and is described as follows:
• An English-speaking human is placed inside a room.
This human does not speak any language other than
English and in particular has no ability to read, speak, or
understand Chinese.
• Inside the room with the human are a set of cards, upon
which are printed Chinese symbols, and a set of
instructions that are written in English.
• A story, in Chinese, is fed into the room through a slot,
along with a set of questions about the story.
“The limits of my language mean the limits of my world.”
-Ludwig Wittgenstein.
The Chinese Room
• By following the instructions that he has, the human is
able to construct answers to the questions from the
cards with Chinese symbols and pass them back out
through the slot to the questioner.
• If the system were set up properly, the answers to the
questions would be sufficient that the questioner would
believe that the room (or the person inside the room)
truly understood the story, the questions, and the
answers it gave.
“The limits of my language mean the limits of my world.”
-Ludwig Wittgenstein.
The Chinese Room
Searle’s argument is now a simple one.
• The man in the room does not understand Chinese. The
pieces of card do not understand Chinese. The room
itself does not understand Chinese, and yet the system
as a whole is able to exhibit properties that lead an
observer to believe that the system (or some part of it)
does understand Chinese
• In other words, running a computer program that
behaves in an intelligent way does not necessarily
produce understanding, consciousness, or real
intelligence.
“The limits of my language mean the limits of my world.”
-Ludwig Wittgenstein.
The Chinese Room
• This argument clearly contrasts with Turing’s view that
a computer system that could fool a human into thinking
it was human too would actually be intelligent.
• One response to Searle’s Chinese Room argument, the
Systems Reply, claims that although the human in the
room does not understand Chinese, the room itself does.
In other words, the combination of the room, the human,
the cards with Chinese characters, and the instructions
form a system that in some sense is capable of
understanding Chinese stories. There have been a great
number of other objections to Searle’s argument, and the
debate continues.
• [Find more other arguments like the Chinese Room]
Human Brain as a Computer
• The Halting Problem and Gِ odel’s incompleteness
theorem tell us that there are some functions that a
computer cannot be programmed to compute, and as a
result, it would seem to be impossible to program a
computer to perform all the computations needed for real
consciousness. This is a difficult argument, and one
potential response to it is to claim that the human brain
is in fact a computer, and that although it must also
be limited by the Halting Problem, it is still capable
of intelligence.
Human Brain as a Computer
• Neural Networks is based on the claim that the human
brain is a computer.
• By combining the processing power of individual
neurons, we are able to produce artificial neural
networks that are capable of solving extremely complex
problems, such as recognizing faces.
• Proponents of strong AI might argue that such
successes are steps along the way to producing an
electronic human being.
• Objectors would point out that this is simply a way to
solve one small set of problems—not only does it not
solve the whole range of problems that humans are
capable of, but it also does not in any way exhibit
anything approaching consciousness.
HAL—Fantasy or Reality?
• In the film 2001: A Space Odyssey. One of the main
characters in the film is HAL, a Heuristically programmed
ALgorithmic computer. In the film, HAL behaves, speaks,
and interacts with humans in much the same way that a
human would, In fact, this humanity is taken to extremes
by the fact that HAL eventually goes mad.
• In the film, HAL played chess, worked out what people
were saying by reading their lips, and engaged in
conversation with other humans.
• How many of these tasks are computers capable of
today? [Games, Natural Language Processing, Machine
Vision
• Finally, the likelihood of a computer becoming insane is
a rather remote one, although it is of course possible that
a malfunction of some kind could cause a computer to
exhibit properties not unlike insanity!
Fantasy or Reality?
• Artificial Intelligence has been widely represented in
other films. The Stephen Spielberg film AI:Artificial
Intelligence is a good example. In this film, a couple buy
a robotic boy to replace their lost son. The audience’s
sympathies are for the boy who feels emotions and is
clearly as intelligent (if not more so) as a human being.
This is strong AI, and while it may be the ultimate goal of
some Artificial Intelligence research, even the most
optimistic proponents of strong AI would agree that it is
not likely to be achieved in the next century
AI in the 21st Century
• Artificial Intelligence is all around us.
• Fuzzy logic, for example, is widely used in washing
machines, cars, and elevator control mechanisms. (Note
that no one would claim that as a result those machines
were intelligent, or anything like it! They are simply using
techniques that enable them to behave in a more
intelligent way than a simpler control mechanism would
allow.)
AI in the 21st Century
• Intelligent agents, are widely used. For example, there
are agents that help us to solve problems while using our
computers and agents that traverse the Internet, helping
us to find documents that might be of interest. The
physical embodiment of agents, robots, are also
becoming more widely used. Robots are used to explore
the oceans and other worlds, being able to travel in
environments inhospitable to humans. It is still not the
case, as was once predicted, that robots are widely used
by households, for example, to carry shopping items or
to play with children, although the AIBO robotic dog
produced by Sony and other similar toys are a step in
this direction.
Part I: Chapter 2: Summary
■ The Chinese Room argument is a thought experiment designed by
John Searle, which is designed to refute strong AI.
■ The computer HAL, as described in the film 2001: A Space Odyssey,
is not strictly possible using today’s technology, but many of its
capabilities are not entirely unrealistic today.
■ The computer program, Deep Blue, beat world chess champion
Garry Kasparov in a six-game chess match in 1997. This feat has
not been repeated, and it does not yet represent the end of human
supremacy at this game.
■ Artificial Intelligence is all around us and is widely used in industry,
computer games, cars, and other devices, as well as being a
valuable tool used in many computer software programs.
Part II: Knowledge
Representation
“If, for a given problem, we have a means of checking a
proposed solution, then we can solve the problem by
testing all possible answers. But this always takes much
too long to be of practical interest. Any device that can
reduce this search may be of value.”
-Marvin Minsky, Steps Toward Artificial Intelligence
Part II: Knowledge
Representation
• The way in which the computer represents a
problem, the variables it uses, and the
operators it applies to those variables can
make the difference between an efficient
algorithm and an algorithm that doesn’t work
at all. This is true of all Artificial Intelligence
problems, and as we see in the following, it is
vital for search.
• The example Contact lens problem
Contact lens problem
• “Imagine that you are looking for a contact
lens that you dropped on a football field.
You will probably use some knowledge
about where you were on the field to help
you look for it. If you spent time in only half
of the field, you do not need to waste time
looking in the other half.”
Contact lens problem
• Now let us suppose that you are having a
computer search the field for the contact
lens, and let us further suppose that the
computer has access to an omniscient
oracle that will answer questions about the
field and can accurately identify whether
the contact lens is in a particular spot.
• Now we must choose a representation for
the computer to use so that it can
formulate the correct questions to ask.
Contact lens problem
Representation 1
• One representation might be to have the
computer divide the field into four equal
squares and ask the oracle for each
square, “Is the lens in this square?”.
• This will identify the location on the field of
the lens but will not really be very helpful
to you because you will still have a large
area to search once you find which quarter
of the field the lens is in.
Contact lens problem
Representation 2
• Another representation might be for the
computer to have a grid containing a
representation of every atom contained in the
field. For each atom, the computer could ask its
oracle, “Is the lens in contact with this atom?”
• This would give a very accurate answer indeed,
but would be an extremely inefficient way of
finding the lens. Even an extremely powerful
computer would take a very long time indeed to
locate the lens.
Contact lens problem
Representation 3
• Perhaps a better representation would be to
divide the field up into a grid where each square
is one foot by one foot and to eliminate all the
squares from the grid that you know are
nowhere near where you were when you lost the
lens. This representation would be much more
helpful.
• In fact, the representations we have
described for the contact lens problem are
all really the same representation, but at
different levels of granularity.
• The more difficult problem is to determine
the data structure that will be used to
represent the problem we are exploring.
• There are a wide range of representations
used in Artificial Intelligence.
• When applying Artificial Intelligence to search
problems, a useful, efficient, and meaningful
representation is essential. In other words, the
representation should be such that the computer
does not waste too much time on pointless
computations, it should be such that the
representation really does relate to the problem
that is being solved, and it should provide a
means by which the computer can actually solve
the problem.
Semantic Nets
• A semantic net is a graph consisting of nodes that are connected
by edges.
• The nodes represent objects.
• The links between nodes represent relationships between those
objects.
• The links are usually labeled to indicate the nature of the
relationship
Semantic Nets
Instances
Semantic Nets
• The links are arrows, meaning that they
have a direction. In this way. It may be that
Fang does chase Fido as well, but this
information is not presented in this
diagram.
• Semantic nets do have limitations, such as
the inability to represent negations: “Fido
is not a cat.”, this kind of fact can be
expressed easily in first-order predicate
logic and can also be managed by rulebased systems.
Download