Research in artificial intelligence (AI) represents an attempt to understand intelligent behaviour (and its prerequisites: perception, language use, and the mental representation of information) by making computers reproduce it. It has historical precedents in the work of Pascal, Leibniz and Babbage, who all devised schemes for intelligent machines that were impractical, given the mechanical components from which it was envisaged that those machines would be constructed. The existence of AI as an independent discipline, however, can be traced to the invention of the digital computer, and, more specifically, to a conference at Dartford College, New Hampshire, in 1956. At that conference, Allen Newell, Cliff Shaw and Herbert Simon (see, e.g. Newell et al. 1957) described the use of heuristic, rule of thumb, procedures for solving problems, and showed how those procedures could be encoded as computer programs. Their work contrasted sharply with immediately preceding attempts to explain intelligent behaviour by modelling the properties of brain cells.
Newell et al.’s work led to the development of more sophisticated programming languages (in particular John McCarthy’s LISP), and their general approach, dubbed semantic information processing by Marvin Minsky, was applied to fields as diverse as visual object recognition, language understanding and chess playing. The advent of larger, faster computers forced AI researchers to confront the problem of whether their programs would scale up so that they could operate in the real world, rather than on small-scale laboratory tasks. Could a program that conversed about wooden blocks on a table top be generalized so that it could talk about poverty in the developing world, for example?
Often the answer was: no. Tricks that worked in a limited range of cases would fail on other cases from the same domain. For example, ideas used in early object recognition programs were later shown to be restricted to objects with flat surfaces. Many aspects of intelligence came, therefore, to be seen as the exercise of domain-specific knowledge, which had to be encoded in detail into computer programs, and separately for each subdomain. This conception of intelligence led to the construction of expert systems. DENDRAL (see Lindsay et al. 1980), which computed the structure of complex organic molecules, and MYCIN (see Shortliffe 1976), which diagnosed serious bacterial infections, were the first such systems, and they remain among the best known. Expert systems have become commercially important, usually in the guise of sophisticated aides memoires for human experts, rather than as replacements for them. One of their principal strengths is their ability to process probabilistic information, which contributes to many kinds of diagnosis.
A different attack on the semantic information processing research of the 1960s came from David Marr (see, in particular, Marr 1982), who argued that AI researchers had failed to provide a computational theory of the tasks their machines were trying to carry out. By a computational theory he meant an account of what outputs those machines were trying to produce from their inputs, and why. Marr wanted to combine evidence from neurophysiology and perceptual psychology with computational techniques from AI, and other parts of computer science, to produce a detailed and explanatory account of human vision. Although he did not wholly succeed before his untimely death in his mid-thirties, his work on the lower levels of the human visual system remains the paradigmatic example of successful research in cognitive science.
Marr’s computational models contained units that were intended to mimic the properties of cells in the visual system. He therefore reintroduced techniques that had been sidelined by Newell et al.’s information processing approach. Connectionism is a more direct descendant of the earlier neural modelling approach. However, the units from which connectionist models are built are not based on specific classes of nerve cells, as those in Marr’s models are.
Marr questioned whether traditional AI could generate explanations of intelligent behaviour, and connectionists asked whether it could model biological, and in particular human, intelligence. Not surprisingly, therefore, many AI researchers in the 1980s embraced the engineering side of the discipline, and focused their attention on its applications. This orientation was also sensible at a time when funding was more readily available for projects with short- to intermediate-term returns. Ideas from AI, albeit not always in a pure form, found their way into commercial expert systems, learning aids, robots, machine vision and image processing systems, and speech and language technology. Indeed, techniques that have their origin in AI are now widespread in many types of software development, particularly of programs that run on personal computers. Object-oriented programming was first introduced in the AI language SMALLTALK. It is now a staple of programming in C++, one of the standard languages for windows applications.
Learning posed another problem for traditional AI. The semantic information processing approach assumed that the analysis of an ability, such as chess playing, should be at an abstract level, independent of the underlying hardware—person or machine—and independent of how the hardware came to have the ability, by learning or by being programmed. The ability to learn is itself an ability that might be modelled in an AI program, and some AI programs were programs that learned. However, a vague unease that learning was not receiving the attention it deserved grew, in some quarters, to the feeling that machines could never be really intelligent unless they could learn for themselves. Furthermore, the kinds of learning modelled in AI programs were limited. For example, concept learning programs were capable only of making slightly more complex concepts out of simpler ones that were programmed into them. For some kinds of machine induction this kind of learning is satisfactory, and it can produce impressive results on large machines. However, it leaves doubts about the limitations of such methods of learning unanswered.
Connectionism, with its techniques of pattern abstraction and generalization, has been seen by many as at least a partial solution to the problems about learning that beset traditional AI. A different approach is based on the analogy between evolution and learning, and on the idea that many types of intelligent behaviour are the product of evolution, rather than of learning in individual animals. Genetic algorithms were originally invented by John Holland (see 1992) who showed, perhaps surprisingly, that computer programs can be evolved by a method that parallels evolution by natural selection. The programs are broken into pieces that can be recombined according to fixed rules (as bits of genetic material are recombined in sexual reproduction). The new programs then attempt to perform the to-be-learned task, and the ones that work best are selected to enter the next round, and to leave offspring of their own. As in evolution, this process must be iterated many times if order is to emerge, in the form of a program that can carry out the required task. The use of genetic algorithms is closely allied to the emerging discipline of artificial life, which is broader in scope and more controversial than AI.
Many of the controversies that surround artificial life are philosophical ones, and they often parallel those generated by AI research. Two main questions have been prompted by work in AI. First, can machines really exhibit intelligence, or can they only simulate it in the way that meteorological computers simulate weather systems? Second, if there were intelligent machines, what moral issues would their existence raise? Potentially, intelligent machines are different from weather-predicting computers, since those computers do not produce rain, whereas robots could act intelligently in the real world. The moral question, like all moral questions raised by scientific advances, must be addressed by society and not specifically by AI researchers.
Alan Garnham
University of Sussex
References
Holland, J.H. (1992) Adaptation in Natural and Artificial Systems: An Introductory Analysis with Applications to Biology, Control, and Artificial Intelligence, 2nd edn, Cambridge, MA.
Lindsay, R., Buchanan, B.G., Feigenbaum, E.A. and Lederberg, J. (1980) Applications of Artificial Intelligence for Chemical Inference: The DENDRAL Project, New York.
Marr, D. (1982) Vision: A Computational Investigation into the Human Representation and Processing of Visual Information, San Francisco, CA.
Newell, A., Shaw, J.C. and Simon, H.A. (1957) ‘Empirical explorations with the Logic Theory Machine: A case study in heuristics’, Proceedings of the Western Joint Computer Conference 15.
Shortliffe, E.H. (1976) ‘A model of inexact reasoning in medicine’, Mathematical Biosciences 23.
Further reading
Boden, M.A. (1987) Artificial Intelligence and Matured Man, 2nd edn, London.
Garnham, A. (1988) Artificial Intelligence: An Introduction, London.
Rich, E. and Knight, K. (1991) Artificial Intelligence, 2nd edn, New York.
