Precursors to AI

Just as there are fictional precursors to AI, so there are artefacts in history - elaborate toys, automata, etc - which, while totally unintelligent, convey by metaphor the possibility of intelligent machines.  Thus, throughout history mechanical principles have been exploited in devices to create the illusion of animation and intelligence.

In all ages, moving models have been constructed to resemble living creatures.  For example, the ancient Greeks, Ethiopians and Chinese built statues and other figures, powered by steam or falling water, to act out sequences of motions.  Pindar noted the animated figures which adorned every public street and which seemed 'to breathe in stone, or move their marble feet' (Olympic Ode, ca 520 B.C.).  Daedalus was said to have devised moving statues worked by quicksilver, which walked in front of the Labyrinth.  And according to Vitruvius, Ctesibius discovered a wealth of pneumatic laws in the third century B.C. and ' devised methods of raising water, automatic contrivances and amusing things of many kinds ... blackbirds singing by means of waterworks and figures that drink and move...'.  Gullible folk often believed that the devices were truly animate, or even worked by divine agency, but there were also robust sceptics: Celsus, for example, writing in the first century A.D., commented scathingly on magic and animals 'not really living but having the appearance of life'.

The fourth century A.D. saw a golden Buddhist statue, set on a carriage and tended by animated models of Taoist monks.  As the carriage moved, the monks circled the Buddha, variously bowing and saluting and throwing incense into a censer.  In the seventh century, boats were constructed with animated figures, and eighth-century Chinese records depict the mechanical figure of a monk which reached out, saying 'Alms! Alms!', and conveying coins from its hands into a satchel.  In 790 A.D. a wooden otter was devised in China, which was said to be able to catch fish, and in 890 a wooden cat, able to catch rats and dancing tiger-flies, was constructed.

Albertus Magnus (1204-72) is said to have manufactured a life-size animated servant.  In one version of the tale, Thomas Aquinas destroyed the automaton when he encountered it in the street, believing it to be the work of the devil.  The creature - made of metal, wood, glass, wax and leather - is said to have been able to talk and open the door for visitors.  Roger Bacon (1214-94) made a speaking head, to the consternation of the pious; and Leonardo da Vinci (1452-1519) constructed an automatic lion in honour of Louis XII: the lion approached Louis, opened its chest with a claw, and pointed to the fleur-de-lis coat of arms of France.  And in the seventeenth century René Descartes Built an automaton, 'ma fille Francine', which a sea captain flung from his ship in superstitious dread.  The eighteenth century saw animated flute players and  talking machines: thus Goethe observed: 'The talking machine of Kempelen is not very loquacious but it pronounces certain childish words very nicely.'

In eighteenth-century Switzerland a number of craftsmen (such as Pierre and Henri-Louis Jaquet-Droz) devised automata that could write, draw pictures and play musical instruments.  For instance, the Scribe (1770), an elegantly dressed figure of a child, could write with a quilled pen that it dipped in ink and the moved over the page.  This remarkable effect was achieved by an elaborate array of precision cams driven by springs.  Similarly, the Draughtsman could produce four drawings, one a portrait of Louis XV.  While the activated cams were changing their positions, the automaton used bellows to blow the dust off the drawing paper.  The Musician - with moving fingers, heaving breast, glancing eyes, etc - was able to play a miniature organ.  The Scribe, Draughtsman and Musician are today held in the Musée d'Art et d'Histoire in Neuchâtel, Switzerland.

The nineteenth century saw a variety of talking machines: for example, Euphonia, a 'bearded Turk' exhibited in the Egyptian Hall in Piccadilly.  The device could ask and answer questions, laugh, whisper and sing.  The movable mouth carried a flexible tongue and an indiarubber palate.  Game-playing automata were also devised in the nineteenth century; and before World War I, the Spanish scientist Leonardo Torres y Quevedo, President of the Academy of Sciences in Madrid, built an electromagnetic automaton which could enable the white king and a rook to mate the black king from any position.  This relatively simple end game was seen as a clever accomplishment in classical mechanics.  A metal base makes contact with the squares of the board to enable the automaton to be informed, by electric currents, of the king's square.  Quevado's son presented the automatic chess player at the 1951 Congress of Cybernetics in Paris.  The celebrated cyberneticist Norbert Wiener was defeated by the machine, and it was jokingly remarked that this was the last victory of classical mechanics over modern cybernetics.

Performing robots were built for the London Radio Exhibition o 1932: the automata could speak, moke cigars and read newspapers.  Alpha, a chromium-plated robot made for the Mullard Valve Company, could tell the time and read aloud daily newspapers (prerecorded each day).  Eric, inspired by R.U.R., opened the exhibition of the Model Engineer Society (1928).  This creature used batteries, two electric motors and a system of belts and pulleys to accomplish a number of limited tasks.  He would rise slowly, bow stiffly, and move his head from side to side.  A loudspeaker in his throat broadcast words from a wireless.  And Elektro, another robot, produced by Westinghouse for the New York World Fair (1939), could achieve twenty-six different movements and respond to spoken commands.  The words were converted into electrical impulses used to operate relays governing an array of motors.  Activated rubber rollers under each foot enabled Elektro to walk.  Sparko, his robot dog, could beg, bark and wag his tail. 

The Festival Plaza of Expo '70 in Osaka witnessed a gigantic robot that carried flashing lights and moved its head.  The device was contrived as part of an overall cybernetic environment, responding to sound and contributing effects of its own.  ONOFF, by contrast, was built from scrap in California to publicise the World Museum at Port Costa.  This robot invited people to insert coins whereupon it produced postcards of itself.  Afterwards people were conducted into the museum to see the large display of toy robots.

The various mechanical devices - animated statues, robots, toys, etc - were designed to simulate living and intelligent systems.  In no sense were such artefacts truly intelligent.  The development of electronics was to change this situation.  For the first time there was the possibility that artificial systems could be configured to embody an intelligent potential.  Mechanical, electrical and electromechanical systems were capable of mimicry and nothing more, but the electronics technologies were to allow the fabrication of artefacts able to perform many activities characteristic of intelligent creatures.  The history of artificial intelligence is in some sense the history of electronic computers.

History of AI

Artificial intelligence is usually regarded as a subclass of computer science, with the implication that there are also other subclasses.  One reason for this convention is that we can date the start of the activities which are the focus of current AI work (see below).  It has also been found convenient to exclude certain computer activities (eg traditional data processing) from the realm of artificial intelligence.  But if an ape or a dolphin were able to compute a complex payroll or actuarial table, we would quickly see such behaviour as evidence for intelligence.  In short, our definition of the AI subclass is usually arbitrary.  An animal that could do differential equations would be deemed intelligent: a similarly skilled computer would not.  And there is also what Hofstadter has called the Tessler Theorem: "Artificial intelligence is whatever computers can't yet do'.  People are reluctant to admit the possibility that artefacts may be intelligent.  But perhaps machines are intelligent if, under their own steam, they can simply do sums.  In fact they are already accomplishing much more.

The first types of calculating devices were the various forms of the abacus, common in ancient China and Japan.  These tools, carrying the familiar rows of beads, were aids for engineers, mathematicians and traders, but in no way could they be considered to be computers: they had no means of storing an internal program of instructions.  Such researchers as Pascal (1647) and Samuel Morland (1666) produced effective mechanical calculators (Samuel Pepys commented on the Morland machine: 'Very pretty but not very useful'), and at about the same time Leibniz devised a calculator that could perform multiplication and division.

Charles Babbage, born in 1792, is often represented as the 'father of modern computing'.  As such he is also a progenitor of artificial intelligence, though this is rarely said.  With Herschel, Babbage created the Royal Astronomical Society in the 1820s, and was obliged to compile reference tables (Babbage: 'I wish to God these calculations had been executed by steam').  He contrived two of the most ambitious calculating machines, celebrating the project in 1822 with a paper to the Society ('Observations on the Application of machinery to the Computation of Mathematical Tables').  At this time he also wrote to Sir Humphrey Davy, president of the Royal Society, proposing that a machine could be developed to replace 'one of the lowest occupations of the human intellect'.

In due course, Babbage developed with mixed success two machines: the Difference Engine and the more ambitious Analytical Engine.  A young Italian military engineer, L F Menabrea, described the latter machine in 1842 (in a paper written in French).  Ada Lovelace, Babbage's co-worker for many years, translated the paper and added her own detailed additions.  These included programs which she had originated.  The original paper was greatly expanded and the power of the Analytical Engine was evident.  Its central importance for the future of electronic computing was that it demonstrated, albeit in mechanical terms, the components that were essential in any general-purpose computer system: input (allowing numbers to be fed into the machine), store (to hold numbers and program instructions), arithmetic unit (to perform the calculations), control unit (to control task performance under the direction of the stored program), and output (to make the result of the processing available to the users).

The tabulator designed by Dr Herman Hollerith, as an effective device for analysing the 1890 American census, was the first computing machine to use non-mechanical processing means.  This approach, allowing an electric current to advance a counter by one, was exploited by International Business Machines (IBM) in its early days.  In 1892, William Burroughs introduced the first commercially available adding machine, but it was not until the early 1940s that it proved possible, following the work of Vannevar Bush with thermionic valves, to use electronic components as elements in digital computing circuits.  The scene was set for the rapid development of the modern electronic digital computer.

Binary operations were introduced by Konrad Zuse in 1935 into the Z1 computer, an entirely mechanical device.  The Z2 used electromechanical relays instead of mechanical switches and employed punched paper tape as input.  The IBM Mark 1 (1943) was also based on electromagnetic relays, and in the same year Colossus 1 began use in Britain to decipher the messages generated by Enigma, the German code system.  Some of the brilliant young men (eg Turing and Michie) involved in code breaking were to become immensely influential in the development of artificial intelligence. 

the 1940s also witnessed the emergence of the Electronic Numerical Integrator and Calculator (ENIAC), designed with 18,000 thermionic valves to compute ballistics tables for guns and missiles.   ENIAC weighed 30 tons and needed to be housed in a room 60ft by 25ft.  The computer was less powerful than any modest micro of the 1980s.  In 1945, John von Neumann began the design of the Electronic Discrete Variable Automatic Computer (EDVAC), and for the first time the notion of stored-program control was incorporated into the design of an electronic digital computer.  Von Neumann was to stimulate one intriguing thread in the AI debate by posting the concept of self-replicating computer systems.  In the same spirit, in 1948, Norbert Wiener advanced the highly influential doctrine that a new science, cybernetics, was equally relevant to self-governing biological and artificial systems.  The ground was being prepared for the analysis of certain types of computer systems in terms of animal psychology and behaviour.

Second-generation computers, based on the transistor rather than the glass valve, were developed in the 1950s.  For example, the Ferranti Mark 1, the Lyons Electronic Office (LEO) and the Ferranti Mercury all became active at this time, and this type of computer technology continued to be commercially dominant throughout the 1960s.  By the early-1970s it was clear, with the capacity to build thousands of effective transistors onto a minuscule silicon chip, that computers would become smaller and more powerful.  Such third-generation technology began to give way to fourth-generation designs in the 1980s, as circuit integration became denser and new programming languages were developed.  The early-1980s also saw the framing of plans for an ambitious fifth generation of computers (see Simons, 1983) in which many aspects of artificial intelligence were intended to play a central part.

One of the most significant threads in the emergence of AI is associated with the name of Alan Turing.  In 1937 he published a seminal paper on 'computable numbers', in which the concept of  the 'universal Turing machine' was launched.  Here it was proposed that a machine could carry out any mathematical procedure, providing the machine was supplied with an adequate instruction table (the equivalent of the modern computer program).  The model that Turing had so skilfully presented was so general that it served to describe all the computers that were to emerge in the decades ahead.  The 'computable numbers' paper is now recognised as one of the most important milestones in the history of computer science.  But more was to come.

After working at Bletchley Park on code breaking during the war, Turing went to the national Physical Laboratory, Teddington, to help to design the Automatic Computing Engine (ACE).  But administrative problems led to delays, and Turing went back to Cambridge in 1947 for a sabbatical.  Here he developed his ideas that an ACE system would be able to model the actions carried out by the human brain, and produced a startlingly prophetic paper on artificial intelligence.  In this paper, 'Computing Machinery and Intelligence', Turing directly addresses the question of whether machines could think, and he observes:

'... I believe that at the end of the century the use of words and general educated opinion will have altered so much that one will be able to speak of machines thinking without expecting to be contradicted.'

He begins the paper with what he calls the 'imitation game', a ploy that is today known as the Turing test.  Here an interrogator is separated from a person (or a machine) under interrogation, and communication is only possible using a teletype.  The idea is that if the human cannot tell, through the interrogation, whether the communication is with another person or a machine, then the machine - if indeed it is a machine giving the answers - may be regarded as intelligent.  Turing was well aware that many people would find absurd the notion that a machine could be intelligent.  So he anticipated some of the objections and answered them (see Objections and Myths under "What is AI".  Turing, with David Champernowne (who later worked on how computers might compose music), also wrote the first chess-playing program, 'Turochamp'.

The administrative difficulties that beset Alan Turing were also to afflict Donald Michie in Edinburgh and other AI researchers.  In 1973 the report by Sir James Lighthill in the UK declared AI work unfruitful and undeserving of government funds.  Work in Great Britain on artificial intelligence received a colossal blow from which it has not yet recovered.  A belated attempt to remedy the situation has been made with the distribution of funds under the 1983/4 Alvey initiative.  But, with regard to the potential of AI, a number of countries - notably the United States and Japan - showed more prescience through the 1970s.

We have seen that the history of artificial intelligence is in some sense the history of computer science.  At the same time it is useful to chart the origins of AI as a subclass discipline.  The term 'artificial intelligence' is usually regarded as having been invented by John McCarthy in 1956, then assistant professor of mathematics at Dartmouth College in Hanover, N.H.  At that time he convened a conference which is seen as the beginning of AI as a separate branch of computer science.  The aim was to bring together serious researchers in the field and to establish effective communication between them.  A number of those who attended - Allen Newell, Herbert Simon, Marvin Minsky and John McCarthy himself - are now universally recognised as leading AI pioneers.

Newell and Simon reported work that they had carried out at the Carnegie Institute of Technology in Pittsburgh (now Carnegie-Mellon University).  They had developed the (now celebrated) theorem-proving Logic Theorist, the first computer program to process symbols rather than numerical quantities.  This became recognised as the first effective AI program.  Working with J C Shaw of the Rand Corporation, Newell and Simon had developed the Information Processing Language (IPL), the first language which enabled computers to process concepts.  The use of IPL to build the Logic Theorist was a major step towards the automation of cognitive thought.

Marvin Minsky, who had worked with Claude Shannon at Bell Laboratories, was to stimulate AI development under project MAC at the Massachusetts Institute of Technology.  He is a co-founder of the MIT AI Group which later became the MIT AI Laboratory.  John McCarthy, another co-founder of the AI Group and now at Stanford University, is the inventor of the Lisp (list processing) language, one of the most favoured AI programming languages.

During the 1970s Edward Feigenbaum, also at Stanford, developed the first expert system, Dendral, used to analyse mass spectrography date.  And another Stanford professor, Terry Winograd, produced a program (called SHRDLU) which was able to manipulate simulated objects shaped like wooden blocks.  This program, much cited in the AI literature, could be told about the simulated blocks and asked to rearrange them.  Dendral and SHRDLU were early examples of programs designed to behave intelligently in particular worlds.  Such programs aim to focus on a particular task and they do not exhibit the generality of response thought by some observers to characterise the truly intelligent system.

It now seems clear that early AI research concentrated unduly on general problem-solving.  The early efforts were largely unsuccessful because of the combinatorial explosion, the fact that exhaustive searches of a problem domain were soon lost in possible paths whose number grew exponentially.  People do not attempt to solve problems in this way.  Instead they rely upon knowledge that is relevant to the problem in question.  It was soon realised that perhaps computers could be programmed to solve well-defined problems in a similar fashion.  This led to a new emphasis on studying how knowledge could be represented in computer systems and inferences drawn from it.  Through the 1970s a main AI theme was the study of knowledge-based systems, often referred to as 'expert systems'.  Other AI interest focused on such things as language translation, game playing and robot behaviour.

In August 1981, Minsky, McCarthy, Newell and others from the Dartmouth conference held a meeting to celebrate the twenty-fifth anniversary of the pioneering encounter.  The meeting was held at the University of British Columbia, Vancouver, during the seventh International Joint Conference on Artificial Intelligence (IJCAI).  More than twenty nations were represented at the international conference, including the USA, USSR, East and West Germany, Sweden, Israel and India.  Over a five-day period, more that 200 papers and panel discussions were presented, and the topics included: expert systems, knowledge representation, inference, search methods, learning, natural language, medical applications and artificial vision.  We may take it as highly significant that many of the papers had a clear psychological or biological relevance.

Today many large companies - IBM, Hewlett-Packard, Digital Equipment Corporation, Tektronix, Fujitsu, Hitachi, etc - have set up AI research laboratories; and important research is being conducted at many institutes and universities (for example, at Stanford, MIT and Carnegie-Mellon in the US, and at Edinburgh and Sussex in the UK).  The Japanese plans for fifth-generation computers require massive funding in AI research and development (see Simons, 1983), and increasing commercial emphasis is being given to the development of particular expert systems for specific purposes (medical diagnosis, crop disease diagnosis, geological prospecting, electronic circuit analysis, chemical synthesis, etc).  Inevitably, military organisations are keen to fund research and development work that will lead to artificially-intelligent weaponry.

Computer systems based on AI techniques are now available as aids for a wide range of professional workers: mathematicians, engineers, doctors, teachers, chemists, geologists, biologists, lawyers, office and factory managers, etc.  There is a growing recognition that AI facilities will become essential to commercial success.  It is important in this context that such facilities need not be expensive.  In fact, expert-systems software is now available for microcomputers.