The hand–eye systems described earlier might be thought of as “robots,” but they could not move about from their fixed base. Up to this time, very little work had been done on mobile robots even though they figured prominently in science fiction. I have already mentioned Grey Walter's “tortoises,” which were early versions of autonomous mobile robots. In the early 1960s researchers at the Johns Hopkins University Applied Physics Laboratory built a mobile robot they called “The Beast.” (See Fig. 12.1.) Controlled by on-board electronics and guided by sonar sensors, photocells, and a “wallplate-feeling” arm, it could wander the white-walled corridors looking for dark-colored power plugs. Upon finding one, and if its batteries were low, it would plug itself in and recharge its batteries. The system is described in a book by Hans Moravec.

Beginning in the mid-1960s, several groups began working on mobile robots. These included the AI Labs at SRI and at Stanford. I'll begin with an extended description of the SRI robot project for it provided the stimulus for the invention and integration of several important AI technologies.

Shakey, the SRI Robot

In November 1963, Charles Rosen, the leader of neural-network research at SRI, wrote a memo in which he proposed development of a mobile “automaton” that would combine the pattern-recognition and memory capabilities of neural networks with higher level AI programs – such as were being developed at MIT, Stanford, CMU, and elsewhere. Rosen had previously attended a summer course at UCLA on LISP given by Bertram Raphael, who was finishing his Ph.D. (on SIR) at MIT.

Difficulties such as those i have just written about rekindled A number of controversies among AI researchers themselves. Frustrated with AI's slowdown, people with different approaches to AI eagerly stepped forward to claim that what AI needed was more of this or that alternative to AI's reigning paradigm – the paradigm John Haugeland called “good-old-fashioned AI” or GOFAI. GOFAI, of course, had as its primary rationale Newell and Simon's belief that a “physical symbol system has the necessary and sufficient means for intelligent action.” But GOFAI seemed to be running out of steam during the 1980s, making it vulnerable to challenges by AI researchers themselves – challenges that had to be taken more seriously than those of Searle, Dreyfus, Penrose, and others outside of the field. In this chapter, I'll describe some of these internal controversies and mention a few of the new paradigms that emerged.

About Logic

Among the pursuers of the GOFAI approach were those who used logical representations and logical reasoning methods – ideas pioneered by John McCarthy. These people were sometimes called “logicists.” (I was among them, having coauthored a 1987 book titled The Logical Foundations of Artificial Intelligence.) Drew McDermott, a professor at Yale University (who received his Ph.D. from MIT), was one of those who began to have doubts about the role of logic in AI. This fact was significant because McDermott himself had been a prominent logicist, but in an influential 1987 paper he concluded that the premise that “… a lot of reasoning can be analyzed as deductive or approximately deductive, is erroneous.”

Throughout the 1980s, while ai was enjoying increased popularity and commercial successes and then suffering funding cuts and a wintry season, its basic research workers produced a significant number of powerful new technical tools and sharpened others. New results unfolded in all of its subfields, including reasoning and representation, machine learning, natural language processing, and computer vision. This work, technically and mathematically deeper than before and strengthened by new connections with statistics and control engineering, helped vitiate some of the criticisms hurled at earlier AI systems and greatly enhanced AI's abilities.

In between the headline-making, milestone ai achievements and the smaller bits of computational intelligence that we find everywhere are impressive AI programs that are used by physicians, scientists, engineers, and business people to help them in (and sometimes automate) their workaday tasks. I call these the “smart tools” of AI. Sometimes these are stand-alone systems, but more often they are integrated into a larger computational framework or into hardware devices. Some work only when called upon to help solve some particular problem, such as disease diagnosis. Some are constantly active, such as online stock-trading systems. I'll not be able to mention all of them since there are far too many, and some are known only to their corporate and government users. But a few examples will serve to illustrate their utility and variety.

In Medicine

Let's start with how AI is being used in medical clinical practice. Beginning as early as the 1980s, AI technology has been an important part of medical systems and devices. In March 2000, a monthly magazine titled Medical Device & Diagnostic Industry published an article claiming that “the medical device industry is seeing an emergence of computer-based intelligent decision support systems (DSSs) and expert systems, the current success of which reflects a maturation of artificial intelligence (AI) technology.” It mentioned several AI-infused devices, including the “Agilent Acute Cardiac Ischemia Time-Insensitive Predictive Instrument…, an intelligent electrocardiagram (ECG) device that predicts the probability of acute cardiac ischemia (ACI), a common form of heart attack,” and the General Electric “MAC 5000 Resting Test System, [incorporating] the Marquette 12SL ECG analysis program, an integrated DSS that uses newly developed digital processing methods and diagnostic program algorithms to interpret and classify ECG waveforms.”

Representing Probabilities in Networks

Much human reasoning is about propositions and quantities that are uncertain. Our beliefs about many things are provisional (that is, subject to change) and qualified (that is, having various levels of confidence). AI systems, too, need to be able to deal with uncertain information. An AI agent's facts, statements, and rules should most appropriately be thought of as provisional and qualified. After all, some of its information is provided by humans and some originates from sensors with limited precision and reliability. Yet, much of the early work in AI ignored the uncertain nature of knowledge. In fact, Marvin Minsky observed that his edited volume of early AI papers contained “no explicit use of probabilistic notions.”

Most AI researchers nowadays, however, acknowledge that much of the knowledge needed by machines needs to be qualified by probability values and that reasoning with this knowledge can therefore most appropriately be done with the tools of probability theory. But just as is the case with logical reasoning, probabilistic reasoning is subject to AI's old nemesis, the combinatorial explosion. Suppose, for example, that an agent's knowledge consists of a set of propositions. Because of possible interdependencies among the propositions, accurate probabilistic reasoning depends on knowing more than just the probability of each of those propositions individually. Instead, probability values for various combinations of the propositions taken together, called “joint probabilities,” are usually required; this leads, in the general case, to impractically large representations and intractable computations.

Earlier AI systems that could deal with uncertainty, such as MYCIN and PROSPECTOR, made simplifying assumptions to ease these representational and computational difficulties.

Clues about what be needed to make machines intelligent are scattered abundantly throughout philosophy, logic, biology, psychology, statistics, and engineering. With gradually increasing intensity, people set about to exploit clues from these areas in their separate quests to automate some aspects of intelligence. I begin my story by describing some of these clues and how they inspired some of the first achievements in artificial intelligence.

From Philosophy and Logic

Although people had reasoned logically for millennia, it was the Greek philosopher Aristotle who first tried to analyze and codify the process. Aristotle identified a type of reasoning he called the syllogism “… in which, certain things being stated, something other than what is stated follows of necessity from their being so.”

Here is a famous example of one kind of syllogism:

1. All humans are mortal. (stated)

2. All Greeks are humans. (stated)

3. All Greeks are mortal. (result)

The beauty (and importance for AI) of Aristotle's contribution has to do with the form of the syllogism. We aren't restricted to talking about humans, Greeks, or mortality. We could just as well be talking about something else – a result made obvious if we rewrite the syllogism using arbitrary symbols in the place of humans, Greeks, and mortal. Rewriting in this way would produce

1. All B's are A. (stated)

2. All C's are B's. (stated)

3. All C's are A. (result)

One can substitute anything one likes for A, B, and C. For example, all athletes are healthy and all soccer players are athletes, and therefore all soccer players are healthy, and so on.

The computer programs i have described so far performed transformation on relatively simple symbol structures, which were all that were required for the mathematical problems, puzzles, and games that these programs dealt with. The main effort was in coming up with and using problem-specific heuristics (such as features to be used in computing the value of a checkers position, for example) to limit the number of transformations of these structures in searches for solutions. As Minsky put it, “The most central idea of the pre-1962 period was that of finding heuristic devices to control the breadth of a trial-and-error search.” In the early 1960s, several Ph.D. research projects, some performed under Minsky's direction at MIT, began to employ more complex symbol structures in programs for performing various intellectual tasks. Because of their rich, articulated content of information about their problem topic, these structures were called semantic representations. As Minsky wrote, “Within the small domain in which each program operates, the performance [of these programs] is not too bad compared with some human activities.… But much more important than what these particular experiments achieve are the methods they use to achieve what they do, for each is a first trial of previously untested ideas.” I'll describe some examples of these sorts of projects and the new methods that they employed.

Solving Geometric Analogy Problems

Thomas G. Evans (1934–) programmed a system that was able to perform well on some standard geometric analogy tests. It was apparently the largest program written up to that time in John McCarthy's new programming language, LISP (which I'll describe later).

In some of the just-preceding chapters, we have seen the important role of data and of machine learning techniques for distilling and using these data. If one single theme has lately begun to unite the several disparate approaches to AI, ranging from logical representations and reasoning in Cyc to decision and estimation by neural networks, it is their dependence on massive amounts of data. Moving beyond toy problems and simple puzzles into real-world problems requires real-world data. In this chapter, I'll explore how the latest systems for natural language processing and computer vision exploit data that are representative of the inputs they must deal with.

Natural Language Processing

The growing need for systems able to deal with written and spoken languages, together with new technical advances, large databases, and increased computational power, has led to improved systems for performing such tasks as summarizing pieces of text, answering queries, and translating languages. In this section, I'll describe some of the technical developments in NLP during the past two or three decades. Impressive as they are though, they have not yet allowed us to realize Terry Winograd's hope back in 1971 that “We will talk to [computer systems] just as we talk to a research assistant, librarian, or secretary, and they will carry out our commands and provide us with the information we ask for.“ Many people say that the problem of realizing such systems is “AI complete,“ in the sense that they must be generally as intelligent as humans, being able to reason and to solve problems as well as humans do those things. In any case, it is probable that such systems, when we finally do have them, will employ some or all of the technology I'll be describing here.

The motivation for much of the computer vision research that i have described during this period was to provide information to guide a robot arm. Because the images that could be analyzed best were of simple objects such as toy blocks, work was concentrated on getting a robot arm to stack and unstack blocks. I'll describe some typical examples of this “hand–eye” research, beginning with a project that did not actually involve an “eye.”

At MIT

A computer-guided mechanical “hand” was developed by Heinrich A. Ernst in 1961 as part of his Electrical Engineering Sc.D. work at MIT. (His advisor was Claude Shannon.) The hand, named MH-1, was a “mechanical servomanipulator [an American Machine and Foundry model 8] adapted for operation by the TX-0 computer.” It used tactile sensors mounted on the hand to guide it because, as Ernst wrote, “organs for vision are too difficult to build at the present time.” The abstract of Ernst's thesis describes some of what the system could do:

The SRI Computer-Based Consultant

As my colleagues and I at SRI cast about for ways to continue our planning and vision research we had been doing under the “Shakey the Robot” project, while satisfying DARPA's interest in militarily relevant applications, we hit upon the problem of equipment maintenance, repair, and training. We pointed out that any technology that could reduce expenditures for these items and lessen the need for utilizing scarce human experts would be extremely important to the military. Furthermore, we said, this need “cannot be satisfied merely by writing more and better manuals. A sophisticated computer system seems to us essential.”

We proposed to begin work on this problem by developing an environment in which an apprentice repair person would be given advice about how to take apart and reassemble a small air compressor. The advice would be generated (as needed by the apprentice) by a planning and execution system that would give suggestions about what to do next. The apprentice could ask for advice, using a predefined small vocabulary, which the system would analyze with state-of-the art speech recognition equipment, and then the system would answer, using a speech generation module. The system would also employ a machine-vision system to help keep track of progress and to observe components pointed to by the apprentice.

Some of ai's recent achievements stand as extraordinary milestones of its progress, and others have insinuated themselves almost invisibly into our daily routines. In between these extremes, AI programs have become important tools in science and commerce. These three categories provide a useful way of organizing the state of AI today. First, I'll look at some of the headline-making systems appearing just before and just after the beginning of the twenty-first century, beginning with AI game-playing programs.

Games

Although getting computers to excel at games, such as chess and checkers, is thought by some to be a somewhat frivolous diversion from more serious work, computer game-playing has served as a laboratory for exploring new AI techniques–especially in heuristic search and in learning. In a previous chapter, for example, I explained how reinforcement learning methods were used to develop a championship-level backgammon program. From the earliest days of AI, people worked on programs to play chess and checkers, and now, mainly by using massive amounts of heuristically guided computation, computers are able to play these and other games better than humans can.

Chess

The big news in 1997 was the defeat of the world chess champion, Garry Kasparov, by IBM's “Deep Blue” chess-playing computer. (See Fig. 32.1.) The first time Kasparov played Deep Blue, in February 1996, Deep Blue won the first game but lost the match. But on May 11, 1997, a hardware-enhanced 1997 version (unofficially nicknamed “Deeper Blue”) won a six-game match (under regular chess tournament time controls) by two wins to one with three draws.

The Logic Theorist and Heuristic Search

Just prior to the Dartmouth workshop, Newell, Shaw, and Simon had programmed a version of LT on a computer at the RAND Corporation called the JOHNNIAC (named in honor of John von Neumann). Later papers described how it proved some of the theorems in symbolic logic that were proved by Russell and Whitehead in Volume I of their classic work, Principia Mathematica. LT worked by performing transformations on Russell and Whitehead's five axioms of propositional logic, represented for the computer by “symbol structures,” until a structure was produced that corresponded to the theorem to be proved. Because there are so many different transformations that could be performed, finding the appropriate ones for proving the given theorem involves what computer science people call a “search process.”

To describe how LT and other symbolic AI programs work, I need to explain first what is meant by a “symbol structure” and what is meant by “transforming” them. In a computer, symbols can be combined in lists, such as (A, 7, Q). Symbols and lists of symbols are the simplest kinds of symbol structures. More complex structures are composed of lists of lists of symbols, such as ((B, 3), (A, 7, Q)), and lists of lists of lists of symbols, and so on. Because such lists of lists can be quite complex, they are called “structures.” Computer programs can be written that transform symbol structures into other symbol structures. For example, with a suitable program the structure “(the sum of seven and five)” could be transformed into the structure “(7 + 5),” which could further be transformed into the symbol “12.”

The Strategic Computing Plan

By the early 1980s expert systems and other AI technologies, such as image and speech understanding and natural language processing, were showing great promise. Also, there was dramatic progress in communications technology, computer networks and architectures, and computer storage and processing technologies. Robert Kahn (1938–; Fig. 23.1), who had become Director of DARPA's Information Processing Techniques Office (IPTO) in 1979, began thinking that DARPA should sponsor a major research and development program that would integrate efforts in all of these areas to create much more powerful computer systems. At the same time, there was concern that the Japanese FGCS program could threaten U.S. leadership in computer technology. With these factors as background, Kahn began planning what would come to be called the “Strategic Computing” (SC) program.

Kahn had been a professor at MIT and an engineer at BBN before he joined DARPA's IPTO as a program manager in late 1972. There he initiated and ran DARPA's internetting program, linking the Arpanet along with the Packet Radio and Packet Satellite Nets to form the first version of today's Internet. He and Vinton Cerf, then at Stanford, collaborated on the development of what was to become the basic architecture of the Internet and its “Transmission Control Protocol” (TCP). (TCP was later modularized and became TCP/IP, with IP standing for Internet Protocol.) Cerf joined DARPA in 1976 and led the internetting program until 1982. For their work, Kahn and Cerf shared the 2004 Turing Award of the Association for Computer Machinery.

There have been naysayers from the erliest days of artificial intelligence. Alan Turing anticipated (and dealt with) some of their objections in his 1950 paper. In this chapter, I'll recount some of the controversies surrounding AI – including some not foreseen by Turing. I'll also describe some formidable technical difficulties confronting the field. By the mid-1980s or so, these difficulties had caused some to be rather dismissive about progress up to that time and pessimistic about the possibility of further progress. For example, in wondering about the need for a special issue of the journal Dœdalus devoted to AI in 1988, the philosopher Hilary Putnam wrote “What's all the fuss about now? Why a whole issue of Dœdalus? Why don't we wait until AI achieves something and then have an issue?”

The attacks and expressions of disappointment from outside the field helped precipitate what some have called an “AI winter.”

Opinions from Various Onlookers

The Mind Is Not a Machine

In the introduction to his edited volume of essays titled Minds and Machines, the philosopher Alan Ross Anderson mentions the following two extreme opinions regarding whether or not the mind is a machine:

Accompanying the technical progress in aptificial intelligence during this period, new conferences and workshops were begun, textbooks were written, and financial support for basic research grew and then waned a bit.

The first large conference devoted exclusively to artificial intelligence was held in Washington, DC, in May 1969. Organized by Donald E. Walker (1928–1993) of the MITRE Corporation and Alistair Holden (1930–1999) of the University of Washington, it was called the International Joint Conference on Artificial Intelligence (IJCAI). It was sponsored by sixteen different technical societies (along with some of their subgroups) from the United States, Europe, and Japan. About 600 people attended the conference, and sixty-three papers were presented by authors from nine different countries. The papers were collected in a proceedings volume, which was made available at the conference to all of the attendees.

Because of the success of this first conference, it was decided to hold a second one in London in 1971. During the early years, organization of the conferences was rather informal, decisions about future conferences being made by a core group of some of the leaders of the field who happened to show up at organizing meetings. At the 1971 meeting in London, I left the room for a moment while people were discussing where and when to hold the next conference.

In september 1948, an interdisciplinary conference was held at the California Institute of Technology (Caltech) in Pasadena, California, on the topics of how the nervous system controls behavior and how the brain might be compared to a computer. It was called the Hixon Symposium on Cerebral Mechanisms in Behavior. Several luminaries attended and gave papers, among them Warren McCulloch, John von Neumann, and Karl Lashley (1890–1958), a prominent psychologist. Lashley gave what some thought was the most important talk at the symposium. He faulted behaviorism for its static view of brain function and claimed that to explain human abilities for planning and language, psychologists would have to begin considering dynamic, hierarchical structures. Lashley's talk laid out the foundations for what would become cognitive science.

The emergence of artificial intelligence as a full-fledged field of research coincided with (and was launched by) three important meetings – one in 1955, one in 1956, and one in 1958. In 1955, a “Session on Learning Machines” was held in conjunction with the 1955 Western Joint Computer Conference in Los Angeles. In 1956, a “Summer Research Project on Artificial Intelligence” was convened at Dartmouth College. And in 1958, a symposium on the “Mechanization of Thought Processes,” was sponsored by the National Physical Laboratory in the United Kingdom.