The introduction to Artificial Intelligence follows. Additional information about this book, along with access to software, is available via http://www.ai.mit.edu/people/phw/Books/

The Intelligent Computer

This book is about the field that has come to be called Artificial Intelligence. In this chapter, you learn how to define Artificial Intelligence, and you learn how the book is arranged. You get a feeling for why Artificial Intelligence is important, both as a branch of engineering and as a kind of science. You learn about some successful applications of Artificial Intelligence. And finally, you learn about criteria you can use to determine whether work in Artificial Intelligence is successful.

The Field and the Book

There are many ways to define the field of Artificial Intelligence. Here is one: Artificial Intelligence is ... the study of the computations that make it possible to perceive, reason, and act. From the perspective of this definition, Artificial Intelligence differs from most of psychology because of the greater emphasis on computation, and Artificial Intelligence differs from most of computer science because of the emphasis on perception, reasoning, and action.

From the perspective of goals, Artificial Intelligence can be viewed as part engineering, part science:

The engineering goal of Artificial Intelligence is to solve real-world problems using Artificial Intelligence as an armamentarium of ideas about representing knowledge, using knowledge, and assembling systems.
The scientific goal of Artificial Intelligence is to determine which ideas about representing knowledge, using knowledge, and assembling systems explain various sorts of intelligence.

This Book Has Three Parts

To make use of Artificial Intelligence, you need a basic understanding of how knowledge can be represented and what methods can make use of that knowledge. Accordingly, in Part I of this book, you learn about basic representations and methods. You also learn, by way of vision and language examples, that the basic representations and methods have a long reach.

Next, because many people consider learning to be the sine qua non of intelligence, you learn, in Part II, about a rich variety of learning methods. Some of these methods involve a great deal of reasoning; others just dig regularity out of data, without any analysis of why the regularity is there.

Finally, in Part III, you focus directly on visual perception and language understanding, learning not only about perception and language per se, but also about ideas that have been a major source of inspiration for people working in other subfields of Artificial Intelligence.

The Long-Term Applications Stagger the Imagination

As the world grows more complex, we must use our material and human resources more efficiently, and to do that, we need high-quality help from computers. Here are a few possibilities:

In farming, computer-controlled robots should control pests, prune trees, and selectively harvest mixed crops.
In manufacturing, computer-controlled robots should do the dangerous and boring assembly, inspection, and maintenance jobs.
In medical care, computers should help practitioners with diagnosis, monitor patients' conditions, manage treatment, and make beds.
In household work, computers should give advice on cooking and shopping, clean the floors, mow the lawn, do the laundry, and perform maintenance chores.
In schools, computers should understand why their students make mistakes, not just react to errors. Computers should act as superbooks, displaying planetary orbits and playing musical scores, thus helping students to understand physics and music.

The Near-Term Applications Involve New Opportunities

Many people are under the false impression that the commercial goal of Artificial Intelligence must be to save money by replacing human workers. But in the commercial world, most people are more enthusiastic about new opportunities than about decreased cost. Moreover, the task of totally replacing a human worker ranges from difficult to impossible because we do not know how to endow computers with all the perception, reasoning, and action abilities that people exhibit.

Nevertheless, because intelligent people and intelligent computers have complementary abilities, people and computers can realize opportunities together that neither can realize alone. Here are some examples:

In business, computers can help us to locate pertinent information, to schedule work, to allocate resources, and to discover salient regularities in databases.
In engineering, computers can help us to develop more effective control strategies, to create better designs, to explain past decisions, and to identify future risks.

Artificial Intelligence Sheds New Light on Traditional Questions

Artificial Intelligence complements the traditional perspectives of psychology, linguistics, and philosophy. Here are several reasons why:

Computer metaphors aid thinking. Work with computers has led to a rich new language for talking about how to do things and how to describe things. Metaphorical and analogical use of the concepts involved enables more powerful thinking about thinking.
Computer models force precision. Implementing a theory uncovers conceptual mistakes and oversights that ordinarily escape even the most meticulous researchers. Major roadblocks often appear that were not recognized as problems at all before the cycle of thinking and experimenting began.
Computer implementations quantify task requirements. Once a program performs a task, upper-bound statements can be made about how much information processing the task requires.
Computer programs exhibit unlimited patience, require no feeding, and do not bite. Moreover, it is usually simple to deprive a computer program of some piece of knowledge to test how important that piece really is. It is almost always impossible to work with animal brains with the same precision.

Note that wanting to make computers be intelligent is not the same as wanting to make computers simulate intelligence. Artificial Intelligence excites people who want to uncover principles that must be exploited by all intelligent information processors, not just by those made of neural tissue instead of electronic circuits. Consequently, there is neither an obsession with mimicking human intelligence nor a prejudice against using methods that seem involved in human intelligence. Instead, there is a new point of view that brings along a new methodology and leads to new theories.

Artificial Intelligence Helps Us to Become More Intelligent

Just as psychological knowledge about human information processing can help to make computers intelligent, theories derived primarily with computers in mind often suggest useful guidance for human thinking. Through Artificial Intelligence research, many representations and methods that people seem to use unconsciously have been crystallized and made easier for people to deploy deliberately.

What Artificial Intelligence Can Do

In this section, you learn about representative systems that were enabled by ideas drawn from Artificial Intelligence. Once you have finished this book, you will be well on your way toward incorporating the ideas of Artificial Intelligence into your own systems.

Intelligent Systems Can Help Experts to Solve Difficult Analysis Problems

During the early days of research in Artificial Intelligence, James R. Slagle showed that computers can work problems in integral calculus at the level of college freshmen. Today, programs can perform certain kinds of mathematical analysis at a much more sophisticated level.

The KAM program, for example, is an expert in nonlinear dynamics, a subject of great interest to scientists who study the equations that govern complex object interactions.

Intelligent Systems Can Help Experts to Design New Devices

The utility of intelligence programs in science and engineering is not limited to sophisticated analysis; many recent programs have begun to work on the synthesis side as well.

For example, a program developed by Karl Ulrich designs simple devices and then looks for cost-cutting opportunities to reduce the number of components. In one experiment, Ulrich's program designed a device that measures an airplane's rate of descent by measuring the rate at which air pressure is increasing.

Intelligent Systems Can Learn from Examples

Most learning programs are either experience oriented or data oriented. The goal of work on experience-oriented learning is to discover how programs can learn the way people usually do---by reasoning about new experiences in the light of commonsense knowledge.

The goal of work on data-oriented learning programs is to develop practical programs that can mine databases for exploitable regularities. Among these data-oriented learning programs, the most well-known is the ID3 system developed by J. Ross Quinlan. ID3 and its descendants have mined thousands of databases, producing identification rules in areas ranging from credit assessment to disease diagnosis.

One typical exercise of the technology, undertaken by Quinlan himself, was directed at a database containing information on patients with hypothyroid disease. This information, provided by the Garvan Institute of Medical Research in Sydney, described each of several thousand cases of thyroid disease in terms of 7 continuous variables (such as the measured level of thyroid-stimulating hormone, TSH), and 16 discrete variables, (such as whether or not the patient had already had thyroid surgery).

From this sea of data, Quinlan's program extracted three straightforward rules for classification of hypothyroid disease:

If     the patient's TSH level is less than 6.05 units 
then   the patient's class is negative 

If     the patient has not had thyroid surgery 
       the patient's TSH level is greater than 6.05 units 
       the patient's FTI level is less than 64.5 units 
then   the patient's class is primary hypothyroid 

If     the patient is not taking thyroxine 
       the patient has not had thyroid surgery 
       the patient's TSH level is greater than 6.05 units 
       the patient's FTI level is greater than 64.5 units 
       the patient's TT4 level is less than 150.5 units 
then   the patient's class is compensated hypothyroid

Evidently, of the seven continuous and 16 discrete variables available, only five are useful in disease classification.

Intelligent Systems Can Provide Answers to English Questions Using both Structured Data and Free Text

As the Voyager 2 spacecraft concluded its 12-year grand tour of the outer planets, it sent back spectacular images of Neptune's moons and rings, much to the delight of journalists and scientists gathered to witness the event. In cooperation with researchers from the Jet Propulsion Laboratory, Boris Katz invited those journalists and scientists to use his START system to ask questions about the Neptune encounter, the Voyager spacecraft, and the Solar system.

To answer straightforward questions, START accessed a variety of tables, including a distance table supplied by the Voyager navigation team and a time-line table supplied by the Voyager sequencing group. Here are a few representative examples:

When was Voyager launched?
The Voyager Spacecraft was launched in Florida on August 20 1977.
At what time will Voyager lose contact with Earth?
The Voyager Spacecraft will lose the contact with Earth around 2014.
How long is the year on Neptune?
Neptune's period of revolution is about 165 Earth years.

START also answered questions by printing out English text drawn from various space publications, as illustrated by the following representative questions and answers. Note that the critical words in the questions---color, talk, and weather---do not appear in the answers:

What color is Neptune?
When you look at a photograph of Neptune, you see the planet's atmosphere and the tops of clouds. Neptune looks bluish because the methane in its atmosphere reflects blue and green frequencies of light. The methane absorbs red and yellow frequencies of sunlight.
How does Voyager talk to Earth?
Commands are sent from Earth to one of two receivers on the Voyager spacecraft. Engineering data and scientific information are sent to Earth over two channels. One channel is on all the time, but both channels are on during encounters with the planets. The radio transmitter on the Voyager transmits with about 25 watts of power, or with about the same power as a refrigerator lightbulb. These faint signals are picked up on Earth by ultrasensitive receivers known as the Deep Space Network.
What's the weather on Neptune?
Neptune receives about a thousand times less sunlight than the Earth. The temperatures there are thus quite low. Typically, the temperature is about -360 degrees Fahrenheit (-218 degrees Celsius).

Artificial Intelligence Is Becoming Less Conspicuous, yet More Essential

The first applications of Artificial Intelligence were mostly motivated by the desire of researchers to demonstrate that Artificial Intelligence is of practical value. Now, as the field is maturing, the development of applications is motivated increasingly by the desire of business people to achieve strategic business goals.

One example of a business-motivated application is the Airport Resource Information System---ARIS---developed by Ascent Technology, Inc., and used by Delta Airlines to help allocate airport gates to arriving flights.

Handling the constraints was not the principal challenge faced by ARIS's developers, however. Other difficult challenges were posed by the need to provide human decision makers with a transparent view of current operations, the need to exchange information with mainframe databases, the need to provide rapid, automatic recovery from hardware failures, and the need to distribute all sorts of information to personnel responsible for baggage, catering, passenger service, crew scheduling, and aircraft maintenance. Such challenges require considerable skill in the art of harnessing Artificial Intelligence ideas with those of other established and emerging technologies.

Criteria for Success

Every field needs criteria for success. To determine if research work in Artificial Intelligence is successful, you should ask three key questions:

Is the task defined clearly?
Is there an implemented procedure performing the defined task? If not, much difficulty may be lying under a rug somewhere.
Is there a set of identifiable regularities or constraints from which the implemented procedure gets its power? If not, the procedure may be an ad hoc toy, capable perhaps of superficially impressive performance on carefully selected examples, but incapable of deeply impressive performance and incapable of helping you to solve any other problem.

To determine if an application of Artificial Intelligence is successful, you need to ask additional questions, such as the following:

Does the application solve a real problem?
Does the application open up a new opportunity?

Throughout this book, you see examples of research and applications-oriented work that satisfy these criteria: all perform clearly defined tasks; all involve implemented procedures; all involve identified regularities or constraints; and some solve real problems or open up new opportunities.

Highlights

Artificial Intelligence is the study of the computations that make it possible to perceive, reason, and act.
The engineering goal of Artificial Intelligence is to solve real-world problems; the scientific goal of Artificial Intelligence is to explain various sorts of intelligence.
Applications of Artificial Intelligence should be judged according to whether there is a well-defined task, an implemented program, and a set of identifiable principles.
Artificial Intelligence can help us to solve difficult, real-world problems, creating new opportunities in business, engineering, and many other application areas.
Artificial Intelligence sheds new light on questions traditionally asked by psychologists, linguists, and philosophers. A few rays of this new light can help us to be more intelligent.