Transcript ttay8_ppt_15b

© 2008 Prentice-Hall, Inc.
Slide 1
Is Artificial Intelligence Real?
© 2008 Prentice-Hall, Inc.
Slide 2
Explain the two basic approaches of artificial intelligence
research.
Describe several hard problems that artificial intelligence
research has not been able to solve yet.
Describe several practical applications of artificial intelligence.
Explain what robots are and give several examples illustrating
what they can—and can’t—do.
© 2008 Prentice-Hall, Inc.
Slide 3

Alan M. Turing was the British mathematician who designed
the world’s first operational electronic digital computer during
the 1940s:




Turing effectively launched the field of artificial intelligence (AI) with
a 1950 paper called “Computing Machinery and Intelligence.”
In 1952 he was professionally and socially devastated when he was
arrested and injected with hormones for violation of British antihomosexuality laws.
The 41-year-old genius apparently committed suicide in 1954, years
before the government made his wartime heroics public.
Four decades after his death, Turing’s work still has relevance to
computer scientists, mathematicians, and philosophers.
© 2008 Prentice-Hall, Inc.
Slide 4
Can Machines Think?

The Turing test:

The test involves two people and a computer.
One person, the interrogator, sits at a terminal and types questions.
The questions can be about anything—math, science, politics, sports,
entertainment, art, human relationships, emotions, etc.
As answers to the questions appear on the screen, the interrogator
attempts to guess whether those answers were typed by the other person
or generated by the computer.




According to Turing, by repeatedly fooling interrogators into thinking it is a
person, a computer can demonstrate intelligent behavior. If it acts
intelligently, it is intelligent.
© 2008 Prentice-Hall, Inc.
Slide 5




Turing did not intend this test to be the only way
to demonstrate machine intelligence; he pointed
out that a machine could fail and still be
intelligent.
Even so, Turing believed that machines would be
able to pass his test by the turn of the century.
So far no computer has come close, in spite of 40
years of AI research.
While some people still cling to the Turing test to
define artificial intelligence, most AI researchers
favor less stringent definitions.
© 2008 Prentice-Hall, Inc.
Slide 6
What Is Artificial Intelligence?

Many computer scientists believe that if a task is easy to perform
with a computer, it can’t be an example of
artificial intelligence.
A more recent textbook definition reflects this point of view:



Artificial intelligence is the study of how to make computers do things
at which, at the moment, people are better. —Elaine Rich, in Artificial
Intelligence
Artificial intelligence is the study of the computations that make it possible
to perceive, reason, and act. —Patrick Henry Winston, in Artificial
Intelligence
© 2008 Prentice-Hall, Inc.
Slide 7


Two common approaches to AI
Use computers to simulate human mental processes

Three inherent problems:




Most people have trouble knowing and describing how they do things.
There are vast differences between the capabilities of the human brain
and those of a computer.
Even the most powerful supercomputers can’t approach the brain’s
ability to perform parallel processing.
The best way to do something with a machine is often very different
from the way people would do it.
© 2008 Prentice-Hall, Inc.
Slide 8

A second approach to AI involved designing intelligent
machines independent of the way people think.




This is a more common approach.
Human intelligence is just one possible kind of intelligence.
A machine’s method for solving a problem might be different from
the human method, but no less intelligent.
Many problems are far too complex to solve all at once.


Break these problems into smaller problems that are easier to solve.
Create programs that can function intelligently when confined to
limited domains.
© 2008 Prentice-Hall, Inc.
Slide 9
Opening Games


One of the first popular domains for AI research was the
checkerboard.
Some AI techniques are still used today in a variety of
applications:

Searching: Looking ahead to the possibilities generated by each
potential move


The staggering number of decision points makes brute-force searching
impractical.
Searching is generally guided by a planned strategy and by rules known
as heuristics.
© 2008 Prentice-Hall, Inc.
Slide 10

Heuristics: Rule of thumb



Heuristics guide us toward judgments that experience tells us are likely
to be true.
For example, in checkers, “Keep checkers in the king’s row as long as
possible.”
Pattern recognition: identifying recurring patterns in input data



The goal of pattern recognition is understanding or categorizing
that input.
The best human chess and checkers players remember thousands of
critical board patterns and know the best strategies for playing when
those patterns or similar patterns appear.
Game-playing programs recognize recurring patterns, too, but not
nearly as well as people do.
© 2008 Prentice-Hall, Inc.
Slide 11



Machine learning: learn from experience


Computer game programs often have trouble identifying situations that are
similar but not identical.
Pattern recognition is probably the single biggest advantage a human game
player has over a computer opponent.
If a move pays off, a learning program is more likely to use that move (or
similar moves) in future games.
Most AI researchers have moved on to more interesting and practical
applications.
© 2008 Prentice-Hall, Inc.
Slide 12
Machine Translation Traps

One early project attempted to create a program that
could translate scientific papers between English and
Russian.




A parsing program (parser) would analyze sentence
structure.
Identify each word by its part of speech.
Look up each word in a translation dictionary and
substitute the appropriate word.
Today, programs like Babel Fish that use machine
learning algorithms help to improve machine
translation.
© 2008 Prentice-Hall, Inc.
Slide 13

Conversation without communication




Joseph Weizenbaum, an MIT professor, designed ELIZA in the
1960s to simulate the role of a therapist in a typed conversation with
a patient.
An ELIZA session can easily deteriorate into nonsense dialog laced
with grammatical errors and inappropriate responses.
ELIZA doesn’t pass the Turing test.
Nonsense and common sense


Part of the problem with natural-language communications is the
massive vocabulary of natural languages.
Natural-language parsing programs have to deal with rules that are
vague, ambiguous, and occasionally contradictory.
© 2008 Prentice-Hall, Inc.
Slide 14





Computers are far more successful dealing with naturallanguage syntax than with semantics—the underlying
meaning of words and phrases.
Computers lack what we call common sense—the wealth
of knowledge and understanding about the world that
people share.
The most successful natural-language applications limit the
domain so that virtually all relevant information can be fed
to
the system.
Natural-language processing has come a long way since
ELIZA’s early conversations.
Computers still can’t pass the Turing test, but they can at
least fool people sometimes.
© 2008 Prentice-Hall, Inc.
Slide 15
Knowledge Bases


AI researchers continue to develop techniques for
representing knowledge in computers.
A Knowledge base contains a system of rules for
determining and changing the relationship among facts in
a database.


Facts stored in a database are rigidly organized into
categories.
Ideas stored in a knowledge base can be reorganized as new
information changes their relationships.
© 2008 Prentice-Hall, Inc.
Slide 16
Artificial Experts

An expert system is designed to replicate the decision-making
process of a human expert.



It requires a knowledge base representing ideas from a specific field
of expertise that is constructed by a user, an expert, or a knowledge
engineer—a specialist who converts the words and actions of experts
into a knowledge base.
Some new expert systems can grow their own knowledge bases
while observing human decision makers doing their jobs.
For most expert systems, the process is still human intensive.
© 2008 Prentice-Hall, Inc.
Slide 17

A complete expert system also includes:



Roles of expert systems



A human interface, which enables the user to interact with the system
An inference engine, which puts the user input together with the
knowledge base, applies logical principles, and produces the
requested expert advice
Aid experts by providing automated data analysis and informed
second opinions
Support non-experts by providing advice based on judgments of one
or more experts
Function within narrow, carefully defined domains
© 2008 Prentice-Hall, Inc.
Slide 18

Expert Systems in Action:




The MYCIN medical expert system outperformed many human
experts in diagnosing diseases.
XCON, one of the most successful expert systems in commercial
use today, has been configuring complex computer systems since
it was developed at Digital Equipment Corporation in 1980.
American Express uses an expert system to automate the process
of checking for fraud and misuses of its no-limit credit card.
Blue Cross/Blue Shield of Virginia uses an expert system to
automate insurance claim processing.
© 2008 Prentice-Hall, Inc.
Slide 19


Boeing Company factory workers use an expert system to locate
the right parts, tools, and techniques for assembling airplane
electrical connectors.
AARON, an automated artist programmed by Harold Cohen—an
artist and professor at the University of California at San Diego—
uses more than 1,000 rules of human anatomy and behavior to
create drawings of people, plants, and abstract objects with a robotic
drawing machine.
© 2008 Prentice-Hall, Inc.
Slide 20
Expert Systems in Perspective

An expert system can perform these tasks:







Help train new employees
Reduce the number of human errors
Take care of routine tasks so workers can focus on more challenging
jobs
Provide expertise when no experts are available
Preserve the knowledge of experts after those experts leave
an organization
Combine the knowledge of several experts
Make knowledge available to more people
© 2008 Prentice-Hall, Inc.
Slide 21


Even with a knowledge base, an expert system isn’t the
machine equivalent of a human expert.
Clearly, knowledge engineers can’t use rules to teach
computers all they need to know to perform useful,
intelligent functions outside narrow domains.
© 2008 Prentice-Hall, Inc.
Slide 22
Image Analysis

Image analysis is the process of identifying objects and shapes
in a photograph, drawing, video, or other visual image.
Image analysis is used for everything from colorizing classic
motion pictures to piloting cruise missiles.
Today’s PCs are capable of running image processing software
with practical applications.



Example: Security programs enable PCs with video cameras to
recognize faces of valid users with a high degree of reliability.
© 2008 Prentice-Hall, Inc.
Slide 23
Optical Character Recognition


Optical character recognition (OCR) software
locates and identifies printed characters embedded in
images, thereby “reading” text.
State-of-the-art OCR programs use several
techniques:




Segmentation of a page into pictures, text blocks, and (eventually)
individual characters
Scaled-down expert system technology for recognizing the
underlying rules that distinguish letters
Context “experts” to help identify ambiguous letters by their context
Learning from actual examples and feedback from a human trainer
© 2008 Prentice-Hall, Inc.
Slide 24


Today’s best programs can achieve up to 99% accuracy; they
can perform even better under optimal circumstances.
OCR technology also can be applied to handwritten text, but
not as reliably.
© 2008 Prentice-Hall, Inc.
Slide 25
Automatic Speech Recognition

Automatic speech recognition systems use pattern recognition techniques
similar to those used by vision and OCR systems, including:





Segmentation of input sound patterns into individual words and phonemes
Expert rules for interpreting sounds
Context “experts” for dealing with ambiguous sounds
Learning from a human trainer
Voice recognition systems with speaker independence—the ability to
recognize speech without being trained—are becoming more common,
making speech recognition practical for more applications.
© 2008 Prentice-Hall, Inc.
Slide 26
Talking Computers


Many computer applications speak like humans by playing
prerecorded digitized speech (along with other digitized
sounds) stored in memory or on disk.
Recorded speech won’t work for applications in which the text
to be spoken is unpredictable—such as a talking word
processor—because all the sounds must be prerecorded.


These types of applications require text-to-speech conversion—the
creation of synthetic speech by converting text files into phonetic
sounds.
With speech synthesis software or hardware, PCs can recite anything
you can type, but with voices that sound artificial and robotic.
© 2008 Prentice-Hall, Inc.
Slide 27
Neural Networks

Neural networks (or neural nets): distributed, parallel
computing systems

Inspired by the structure of the human brain

Uses a network of a few thousand simpler processors called
neurons

Not programmed in the usual way—they’re trained


Learns patterns by trial and error, just as the brain does
Optimistic researchers hope that neural networks may someday
provide hearing for the deaf and eyesight for the blind.
© 2008 Prentice-Hall, Inc.
Slide 28
What Is a Robot?


A robot is a computer-controlled machine designed to perform
specific manual tasks.
A robot’s central processor might be a microprocessor
embedded in the robot’s shell, or it might be a supervisory
computer that controls the robot from a distance.


The processor is functionally identical to the processor found
in a computer.
The most important hardware differences between robots and
other computers are the input and output peripherals.
© 2008 Prentice-Hall, Inc.
Slide 29
Steel-Collar Workers

From a management point of view, robots offer several
advantages:




Robots save labor costs.
Robots can work 24 hours a day, 365 days a year, without vacations,
strikes, sick leave, or coffee breaks.
Robots can also improve quality and increase productivity.
 They’re especially effective at doing repetitive jobs.
Robots are ideal for jobs that are dangerous, uncomfortable, or impossible
for human workers.
© 2008 Prentice-Hall, Inc.
Slide 30

As it matures, AI technology finds its way out of the
research lab and into the marketplace.
A growing number of programs and products incorporate
pattern recognition, expert systems, and other AI
techniques.
In the near future we’re likely to see more products with
embedded AI, including:





Intelligent word processors that can help writers turn rough
drafts into
polished prose
Smart appliances that can recognize and obey their owners’
spoken commands
Vehicles that can perform their own diagnostics and, in many
cases, repairs
© 2008 Prentice-Hall, Inc.
Slide 31




Some AI researchers try to simulate human intelligent
behavior, but most try to design intelligent machines
independent of the way people think.
Successful AI research generally involves working on
problems with limited domains rather than trying to
tackle large, open-ended problems.
AI programs employ a variety of techniques, including
searching, heuristics, pattern recognition, and machine
learning, to achieve their goals.
AI researchers have developed a variety of schemes for
representing knowledge in computers.
© 2008 Prentice-Hall, Inc.
Slide 32




We’ll also see more distributed intelligence—AI concepts
applied to networks rather than to individual computers.
A robot is a computer-controlled machine designed to perform
specific manual tasks.
As robot technology advances, artificial workers will do more
traditional human jobs.
In spite of the numerous difficulties AI researchers encounter
when trying to produce truly intelligent machines, many experts
believe that people will eventually create artificial beings that are
more intelligent than their creators—a prospect with staggering
implications.
© 2008 Prentice-Hall, Inc.
Slide 33
Slide
34