Introduction
The invention of computers--based on the work
of Alan Turing in the 1930s and John von Neumann
in the 1950s--quickly gave rise to the notion
of artificial intelligence, or AI, the claim that
such nonhuman machines can exhibit intelligence
because they mimic (or so its proponents claim)
what humans do when they do things we regard as
being evidence of intelligence.
From about the late 1960s to the middle of the
1980s there was a great deal of excitement and
debate among philosophers, psychologists, learning
theorists, and others concerning the possibility
and status of AI. Mostly there were AI champions
and AI detractors, with little middle ground.
That controversy seems to have cooled of late,
but new developments in the computer-engineering
field may now take us past those earlier debates.
(Agre & Chapman 1987)
Research in the provocatively named field of artificial
intelligence (AI) evokes both spirited and divisive
arguments from friends and foes alike. The very
concept of a "thinking machine" has
provided fodder for the mills of philosophers,
science fiction writers, and other thinkers of
deep thoughts. Some postulate that it will lead
to a frightening future in which superhuman machines
rule the earth with humans as their slaves, while
others foresee utopian societies supported by
mechanical marvels beyond present ken. Cultural
icons such as Lieutenant Commander Data, the superhuman
android of Star Trek:
The Next Generation, show a popular willingness
to accept intelligent machines as realistic possibilities
in a technologically advanced future. (Albus 1996)
However, superhuman artificial intelligence is
far from the current state of the art and probably
beyond the range of projection for even the most
optimistic AI researcher. This seeming lack of
success has led many to think of the field of
artificial intelligence as an overhyped failure--yesterday's
news. Where, after all, are even simple robots
to help vacuum the house or load the dishwasher,
let alone the Lieutenant Commander Datas. It therefore
may amaze the reader, particularly in light of
critical articles, to learn that the field of
artificial intelligence has actually been a significant
commercial success.
In fact, according to a 1994 report issued by
the U.S. Department of Commerce, (Critical technology
assessment…1994) the world market for AI
products in 1993 was estimated to be over $900
million!
The reason for this is, in part, that the fruits
of AI have not been intelligent systems that carry
out complex tasks independently. Instead, AI research
to date has primarily resulted in small improvements
to existing systems or relatively simple applications
that interact with humans in narrow technical
domains. While selling well, these products certainly
don't come close to challenging human dominance,
even in today's highly computerized and networked
society.
Some systems that have grown out of AI technology
might surprise you. For example, at tax time many
of you were probably sitting in front of your
home computer running packages such as TURBOTAX,
MACINTAX, and other "rule-based" tax-preparation
software. Fifteen years ago, that very technology
sparked a major revolution in the field of AI,
resulting in some of the first commercial successes
of a burgeoning applied-research area. Perhaps
on the same machine you might be writing your
own articles, using GRAMMATIK or other grammar-checking
programs. (Araujo & Grupen 1996) These grew
out of technology in the AI sub field of "natural
language processing," a research area "proven"
to be impossible back in the late 1960s. Other
examples range from computer chips in Japanese
cameras and TVs that use a technique ironically
called fuzzy logic that improves image quality
and reduces vibration, to an industrial-scale
"expert system" that plans the loading
and unloading of cargo ships in Singapore.
If you weren't aware of this, you are not alone.
Rarely has the hype and controversy surrounding
an entire research discipline been as overwhelming
as it has for the AI field. (AI: The Tumultuous
History…1993) The history of AI includes
raised expectations and dashed hopes, small successes
sold as major innovations, significant research
progress taking place quietly in an era of funding
cuts, and an emerging technology that may play
a major role in helping shape the way people interact
with the information overload in our current data-rich
society.
Where Artificial Intelligence Has Been
Roughly speaking, AI is more than fifty years
old--the field as a coherent area of research
usually being dated from the 1956 Dartmouth conference.
That summer long conference gathered ten young
researchers united by a common dream: to use the
newly designed electrical computer to model the
ways that humans think. They started from a relatively
simple-sounding hypothesis: that the mechanisms
of human thought could be precisely modeled and
simulated on a digital computer. This hypothesis
forms what is, essentially, the technological
foundation on which AI is based.
In that day and age, such an endeavor was incredibly
ambitious. Now, surrounded by computers, we often
forget what the machines of forty years ago looked
like. In those early days, entering a program
into difficult-to-use, noisy teletypes, which
were interfaced with large, snail-paced computers,
largely performed AI. After starting the program
(assuming one had access to one of the few interactive,
as opposed to batch, machines), one would head
off to lunch, hoping for a run of the program
before the computer crashed. In those days, 8K
of core memory was considered major computing
memory, and a 16K disk was sometimes available
for the main memory. In fact, anecdote has it
that some of the runs of Herb Simon's earliest
AI systems used his family and students to simulate
the computations--it was faster than using the
computer! (Beer 1990)
Within a few years, however, AI seemed really
to take off. Early versions of many ambitious
programs seemed to do well, and the thinking was
that the systems would progress at the same pace.
In fact, the flush of success of the young field
led many of the early researchers to believe that
progress would continue at this pace and that
intelligent machines would be achieved in their
lifetimes. A checkers-playing program beat human
opponents, so could chess be far beyond? Translating
sentences from codes (like those developed for
military use during World War II and the Korean
War) into human-understandable words by computer
was possible, so could translation from one human
language to another be too much harder? Learning
to identify some kinds of patterns from others
worked in certain cases, so could other kinds
of learning be much different? (Beer 1995)
Unfortunately, the answers to all of these questions
turned out to be yes. For technical reasons, chess
is much harder than checkers to program.
Translating human languages turns out to have
very different complexities from those encountered
in decoding messages. The learning algorithms
were shown to be severely limited in how far they
could go. In short, the early successes were misleading,
and the expectations they raised were not fulfilled.
At this point, things started getting complicated.
Waiting in the wings and watching carefully were
a number of people who were sure that this new
technology would be a failure. Both philosophers
and computer scientists were sure that getting
computers to "think" was impossible,
and they confused the early difficulties with
fundamental limits.
The problems were magnified tremendously by
the naysayers, who were using arguments about
the theoretical limits (a current example of such
theoretical arguments is Searle's Chinese room
argument) to describe failures of current technology.
In short, those waiting to call the field a flop
felt sure they were seeing evidence to that effect.
(Beer 1997)
One can dwell at length on the early failures;
there were plenty to go around. But as should
have been clear to AI's critics, these failures
were not tragic. In fact, often they were extremely
informative. This should come as no surprise;
after all, this is how science works. Past failures
coupled with new technologies led to many of the
major advances in science's history. It was the
failure of alchemy coupled with better measurement
techniques that led to elemental chemistry; the
newly rediscovered telescope coupled with the
failures of epicycles led to acceptance of the
heliocentric model of the solar system, and so
forth.
In AI, these breakthroughs were less dramatic,
but they were occurring. The exponential improvements
in computing technology (doubling in speed and
memory size every few years), coupled with increasingly
powerful programming languages, made it easier
for AI scientists to experiment with new approaches
and more ambitious models. In addition, each "failure"
added more information for the next project to
build on. Science progressed and much was learned,
often to the chagrin of AI's critics. (Boden 1996)
Critics vs. Technology—The Example Of
Computer Chess
A good example of how this progression occurred
is in the area of chess-playing programs. By the
end of the 1950s, computers were playing a pretty
good game of checkers. A famous checker-playing
program written by Arthur Samuel (who was not
a very good player) had actually beaten him by
the late 1950s, and in 1962 it beat Robert Nealy,
a well-respected player (ex-Connecticut state
champion). Chess seemed just around the corner,
and claims that "in ten years, the best player
in the world will be a machine" were heard.
It turns out, however, that checkers can be played
fairly well using a simple strategy called "minimaxing."
Each move in checkers has at most a few responses,
and searching for the best move doesn't require
examining too many possibilities. The complications
of chess, on the other hand, grow very quickly.
Consider:
There are twenty moves the first player can
make, each followed by twenty possible responses.
Thus, after each player has moved once, there
are about four hundred possible chess boards that
could result. The first player then moves again
(another twenty or more possibilities), and thus
there are now 400 x 20 = 8,000 possible ways the
game could have gone. This sort of multiplying
goes on for a long time--in fact, the calculation
for how many total possibilities exist in a game
of chess was estimated to be about 10[120]. For
even today's fastest supercomputer to examine
all of the possibilities would take over 10[100]
years--well beyond the probable death of the universe.
(Brooks 1991)
Given the complexity of chess, it's hardly surprising
that early programs didn't do very well. In fact,
it wasn't until 1967 that a chess program (written
by Richard Greenblatt, an MIT graduate student)
competed successfully against human players at
the lowest tournament levels. Greenblatt used
a number of techniques being explored by AI scientists
and tailored them for chess play. His program
played chess at a rating reported to be between
1400 and 1500, below (but not far below) the average
rating for players in the U.S. Chess Federation
-- and certainly better than most human neophytes.
In 1965, while researchers were still trying
to figure out how to get the chess-playing programs
to overcome the combinatoric problems (i.e., the
plethora of possible moves), Hubert Dreyfuss,
one of the most outspoken critics of AI to this
day, produced a report for the RAND Corporation
that trashed AI. He argued, both philosophically
and computationally, that computers could never
overcome combinatoric problems. In fact, he stated
categorically that no computer could ever play
chess at the amateur level and certainly that
no computer could beat him at chess. A couple
of years later, he attempted to prove this by
playing against Greenblatt's program. He lost.
(Dorffner 1997a)
Now, nearly thirty years later, the world's best
chess player is still not a machine. However,
today there are a number of computer programs
playing at the master level, and a few that are
breaking into the rank of grand master. In a recent
official long game, a computer beat a player ranked
as the thirtieth best in the United States. Reportedly,
in an "unofficial" short game recently,
a chess program running on a supercomputer beat
Gary Kasparov, arguably the best human player
in the world. Most AI researchers believe that
it is only a matter of a few years until computer
chess programs can beat players of Kasparov's
caliber in official long matches.
What was happening in chess was happening (although
somewhat less dramatically) in many other parts
of the field. It turned out that most of the problems
being looked at by AI researchers suffered from
combinatoric problems, just as chess had. As in
chess, coming up with both better machines and,
more important, better techniques for "pruning"
the large number of possibilities, led to significant
successes in practice. In fact, in the late 1970s
and early '80s, AI was ready to come out of the
laboratories, and it would have a great impact
on the business (and military) world. (Dorffner
1997b)
The Breakthrough: What’s Hard Is Simple
What was the realization that led to the first
successes of AI technology? The intuition was
very simple. Many of the first problems that AI
looked at were ones that seemed easy. If one wanted
to try to get the computer to read books, why
not start with children's stories--after all,
they're the easiest, right? If one wanted to study
problem solving, try basic logic puzzles like
those given on low-level intelligence tests. In
short, it seems obvious that to try to develop
intelligent programs, one should first attack
the problems that humans find easy. The real breakthrough
in AI was the realization that this was just plain
wrong! (Franklin 1885)
In fact, it turns out that many tasks that humans
find easy require having a broad knowledge of
many different things. To see this, consider the
following example from the work of Roger Schank
in the mid-1970s. If a human (or AI program) reads
this simple story: "John went to a restaurant.
He ordered lobster. He ate and left." and
is asked "What did John eat?," the answer
should be "Lobster." However, the story
never says that. Rather, your knowledge about
eating in restaurants tells you that you eat what
you order. Similarly, you could figure out that
John most likely used a fork, that the meal was
probably on the expensive side, that he probably
wore one of those silly little bibs, and so forth.
Moreover, if I mentioned "Mary" was
with him in the restaurant, you'd think about
social relations, dating customs, and lots more.
In fact, to understand simple stories like this,
you must bring to bear tremendous amounts of very
broad knowledge. (Funes & Pollack 1997)
Now consider the following "story" from
the manual for a personal computer hard-disk drive.
If this equipment does cause interference to radio
or TV reception, which can be determined by turning
the equipment on and off, the user is encouraged
to try to correct the interference by one or more
of the following measures: reorient the receiving
antenna, relocate the computer with respect to
the antenna, plug the computer into a different
outlet so that the computer and receiver are on
different branch circuits.
For a human, this story seems much harder to understand
than the one about the lobster. However, if you
think about it, you'll realize that if your computer
was given a fairly narrow amount of knowledge
(about antennas, circuits, TVs, etc.) it would
be able to recognize most of the important aspects
of this story. No broad knowledge is needed to
handle this. Rather, "domain specific"
information about a very narrow aspect of the
world is sufficient. In fact, this is much easier
information to encode. Thus, developing a system
that is an "expert" in hard disks is
actually much easier than developing one that
can handle simple children's stories. (Johnson
1989)
Many of these narrow technical domains can be
of great use. Recognizing what disease someone
has from a set of symptoms, deciding where to
drill for oil based on core samples, figuring
out what machine can be used to make a mechanical
part, configuring a computer system, troubleshooting
a diesel locomotive, and hundreds of other problems
require narrow knowledge about a specific domain.
Building a system that has an expertise in a specific
area proves, in many ways, to be easier than making
one for "simple" tasks.
Spurred by this realization, AI researchers developed
programming technologies, known as rule-based
systems or blackboard architectures, in the mid-to
late 1970s. By the early 1980s, the term expert
system came to be used to describe a program that
could reason (or more often, help a human reason)
through a specific hard problem. The rule bases
could be embedded as parts of larger programs
(such as control systems, decision support tools,
CAD/CAM tools, and others) or used by themselves
with humans providing inputs and outputs. As more
and more industries and government agencies began
to realize the potential for these systems, small
AI companies were started, major companies started
AI laboratories, and the AI boom of the 1980s
was on. (Langton 1989)
The Artificial Intelligence Boom
The 1980s was an exciting time to be an AI scientist.
One didn't know upon waking up in the morning
if he would hear a news story about how AI was
the magic bullet that would solve all the world's
woes or a critical piece about why expert systems
weren't really "intelligent," written
by the critics who had crowed over AI's failures
(and were eating crow over its perceived successes).
Attendance at AI conferences swelled from hundreds
in the late 1970s to thousands in the early 1980s.
Any company that could build rule bases and afford
some basic equipment declared itself an AI expert.
The early days of AI entrepreneurship were very
similar to those of biotechnology or other high-technology
industries. Many companies started, but most were
unsuccessful. The few successes had to change
products and techniques based on market forces;
many look very different than what their founders
expected. (Newell 1982)
However, AI technology, as it has matured and
transitioned, has also become easier to use and
more integrated with the rest of the software
environment. So prevalent is this technology today
that virtually every major U.S. high-technology
firm employs some people trained in AI techniques--in
fact, according to the Commerce Department report
cited earlier, an estimated 70-80 percent of Fortune
500 companies use AI technology to some degree.
Strangely, despite the economic success of this
technology, its long-run effect was to give AI
something of a black eye in the marketplace. There
are many reasons for this, but basically they
boil down to a striking phenomenon: AI is a victim
of its own success. So fast was the transition
of this technology into the marketplace that in
only ten years the necessary technology fell in
cost and complexity by about one to two orders
of magnitude.
Ten years ago, a special computer costing tens
of thousands of dollars was needed to develop
expert systems, but now they can be developed
on generic workstations or even personal computers
costing only a thousand dollars or so. Where the
development environment for an expert system (called
a shell) used to cost $20,000, today one can be
bought for the price of the manuals. (In fact,
numerous shells are available free on the Internet.)
Thus, having the ability to build expert systems
is no longer a high-cost investment; now anyone
technically competent can do it, and do it cheaply.
(Pfeifer & Scheier 1998)
The Artificial Intelligence ‘Bust’?
Unfortunately, there was a negative consequence
of this drop in cost that emerged in the mid-to
late 1980s. Because a lot of money was being invested
in AI and anyone could enter the field, a great
many people did so. Unfortunately, many of these
newcomers had not learned the historical lessons
of AI. Its rapid progress on some problems caused
many to feel this would be easy to extend to other
problems--if AI could handle hard tasks, certainly
it could handle "easier" tasks such
as reading newspapers, translating languages,
playing games, and similar tasks. Even worse,
people with little concept of the combinatorics
of AI tasks would underbid on big development
projects and then be unable to deliver two or
three years later. Thus, many who joined late,
unaware of the field's history, made many of the
same mistakes as had been made in the earliest
days of the discipline. (Steels 1994)
Moreover, it turned out that many of the best
expert systems didn't function by themselves.
Instead of being stand-alone systems that dispatched
wisdom, expert systems turned out to be most useful
when hidden behind larger applications. Take,
for example, the DART (Dynamic Analysis and Replanning
Tool) system developed in the early 1990s by Bolt,
Baranek, and Newman. DART is a military transport
planning program that was used by the U.S. military
in Desert Shield and Desert Storm. It works by
providing a graphical interface in which humans
enter information about what materiel is going
where and when. The system uses its knowledge
to project delivery dates and to recognize possible
problems in meeting those dates.
When a problem is found, DART does not fix it.
Rather, it reports the information to the human
user and asks what to do. Thus, the expertise
in this system is not in making the "intelligent"
decisions about what to do but rather in taking
into account fairly prosaic low-level details
and managing them for the user. In fact, this
is true of most successful expert systems--the
system functions more like a well-trained assistant
than like an expert.
This is not a condemnation, however. DART is
credited by the personnel at the U.S. Advanced
Research Projects Agency (ARPA), the main government
funder of AI research, as having "more than
offset all the money that [ARPA] had funneled
into AI research in the last 30 years." (Vaario
& Ohsuga 1997)
Unfortunately, despite the success of programs
like DART, their interactive nature helped feed
into a subtle negative perception that expert
systems were not successful. Basically, once the
tools reached a certain point of maturity, it
became relatively easy to see how these systems
worked. Understanding that the programs were only
manipulating simple facts or recognizing simple
patterns, people realized the programs were not
"intelligent" at all, that humans were
providing most of the "thinking" and
the AI systems were just managing details. This
gave the naysayers more ammunition--expert systems
clearly were not intelligent by any obvious definition.
Given the hype over these systems, many people
were disappointed to find out that they were just
relatively straightforward computer programs.
In short, what was an industrial success proved
to be insufficient to refute our critics' condemnations--they
won't be satisfied until we build Mr. Data.(Tani
& Nolfi 1998)
The Debate Goes On
Unfortunately, even great strides in information
technology will not bring a "smart"
computer. As the technology reaches fruition,
again the AI field will be accused of just adding
technology, not "developing intelligence."
I suspect that each time that AI surpasses our
current expectations and achieves results changing
the way we live, work, and interact with computers,
the ever-present critics will be fight with us.
In fact, probably no level of success will still
the voices that accuse us of inflated claims,
deflating our successes and denying, to the very
end, the very possibility of artificial intelligence.
|