artificial intelligence

(AI)

Igor Alexander
Margaret Boden
Ron Chrisley

Igor Alexander

The process of designing machines with abilities modelled on human thought. While this mostly involves writing computer programs with human-like characteristics, it has implications for the design of robots and raises philosophical questions about machine–human comparisons.

1. Origins and ambitions
2. Knowledge, logic, and learning
3. Evolution, agents, and brain–mind comparisons

1. Origins and ambitions

Artificial intelligence may be said to have begun in 1950 when Claude Shannon of the Bell Telephone Laboratories in the United States wrote an ingenious program that was to be the forerunner of all chess-playing machines. This work drastically changed the accepted perception of stored-program computers which, since their birth in 1947, had been seen just as automatic calculating machines. Shannon's program added the promise of automated intelligent action to the actuality of automated calculation.

In Shannon's program the programmer stores in the computer the value of important features of board positions. A 'checkmate' being a winning position would have the highest value and the capture of more or less important pieces would be given relatively lower values. So, say that the computer is to take the next move, it would (by being programmed to follow the rules of the game) work out all the possible moves that the opponent might take. It could then work out which moves are available to itself at the next playing period and so on for several periods ahead in the search for a winning path through this 'tree' of possibilities. Sadly, the amount of computation needed to evaluate board positions grows prodigiously the further ahead the computer is meant to look. This process of searching through a large number of options became central in AI programs throughout the 1960s and the early 1970s. Other intelligent tasks besides game playing came under scrutiny: general problem solving, the control of robots, computer vision, speech processing, and the understanding of natural language. Solving general problems requires searches that are similar to those in the playing of board games. For example, to work out how to get from an address in London to the Artificial Intelligence laboratory at the University of Edinburgh, the problem can be represented as a search among subgoals (e.g. get to Edinburgh airport) and the use of 'means' such as airlines, taxis, or railways. The paths through the scheme are evaluated in terms of the reduction of cost and/or the reduction of time to the user. Robot control is similar. The physical rearrangement of objects in a space has to follow a strategy that involves the most efficient path between a current arrangement and the desired one, via several intermediate ones.

Computer vision and the recognition of speech required the programmer to determine that some features of the sensory input generated as signals from a video camera or a microphone are important. For example, for face recognition, the program has to identify the central positions of eyes, nose, and mouth and then measure the size of these objects and the distances between them. These measurements for a collection of faces are stored in a database, each together with the identity of the face. So were a face in the known set to be presented to the camera, finding the closest fit to the measurements stored in the database could identify it. Similarly the features of voices and the sound of words could be stored in databases for the purpose of eventual recognition.

Perhaps the most ambitious target for AI designers was the extraction of meaning from language. This goes beyond speech recognition and sentences could be presented in their written form. The difficulty is for the programmer to find rules that distinguish between sentences such as 'he broke the window with a rock' and 'he broke the window with a curtain'. This required a storage of long lists of words indicating whether they were instruments (rock) or embellishments (curtain) so that the correct meaning could be ascribed to them as they appear in a sentence.

However, early enthusiasm that AI computers could perform tasks comparable to those of humans were to be curtailed by the mid-1970s when poignant shortcomings emerged because the techniques used suffered from serious limitations. In 1971, the British mathematician Sir James Lighthill advised the major science-funding agency in the United Kingdom that AI was suffering from something he called the 'combinatorial explosion' which has been mentioned above in the chess-playing example. Every time the computer needs to look a further step ahead, the number of moves to be evaluated is that of the previous level multiplied by a large amount. In 1980, US philosopher John Searle levelled a second criticism at those who had claimed that they had enabled computers to understand natural language. Through his celebrated 'Chinese Room' argument he pointed out that the computer, by stubbornly following rules, was like a non-Chinese speaker using a massive set of rules to match questions expressed in Chinese symbols about a story also written in Chinese symbols. Given the time to examine many rules, the non-Chinese speaker could find the correct answers in Chinese symbols, without there being any understanding in the procedure. According to Searle, understanding requires a feeling of 'aboutness' for words and phrases which computers do not have. Also a third difficulty began to emerge: artificial intelligence depended too heavily on programmers having to work out in detail how to specify intelligent tasks. In pattern recognition, for example, ideas about how to recognize faces, scenes, and sounds turned out to be inadequate, particularly with respect to human performance.

2. Knowledge, logic, and learning

These censures had a healthy effect on AI. The 1980s saw a maturing of the field through the appearance of new methodologies dubbed knowledge-based systems, expert systems, and artificial neural networks (or connectionism). Effort in knowledge-based systems used formal logic to greater effect than before. The application of the logical rules of inheritance and resolution made more efficient use of knowledge stored in databases. For example, 'Socrates is a man' and 'All men are mortals' could lead to the knowledge that 'Socrates is mortal' by logical inference rather than by explicit storage, thus easing the problem of holding vast amounts of data in databases.

Expert systems was the name given to applications of AI which sought to transfer human expertise into knowledge bases so as to make such knowledge widely available to non-experts. This employed a 'knowledge engineer' who elicited knowledge from the expert and structured it appropriately for inclusion in a database. Facts and rules were clearly distinguished to enable them to be logically manipulated. Typical applications are in engineering design and fault finding, medical diagnosis and advice, and financial advice.

The aim of artificial neural network studies is to simulate mechanisms which, in the brain, are responsible for mind and intelligence. An artificial neuron learns to respond or not to respond ('fire' or not) to a pattern of signals from other neurons to which it is connected. A multi-layered network of such devices can learn (by automatically adjusting the strengths of the interconnections in the network) to classify patterns by learning to extract increasingly telling features of such patterns as data progresses through the layers. The presence of learning overcomes some of the difficulties previously due to the programmer having to decide exactly how to recognize complex visual and speech patterns. Also, a totally different class of artificial neural networks may be used to store and retrieve knowledge. Known as dynamic neural networks, such systems rely on the inputs of neurons being connected to the outputs of other neurons. This allows the net to be taught to keep a pattern of firing activity stable at its outputs. It can also learn to store sequences of patterns. These stable states or sequences are the stored knowledge of the network which may be retrieved in response to some starting state or a set of inputs also connected to the neurons in the net. So in terms of pattern recognition, not only can these networks learn to label patterns, but also 'know' what things look like in terms of neural firing patterns.

3. Evolution, agents, and brain–mind comparisons

Despite the above two major phases in the history of artificial intelligence, the subject is still developing, particularly in three domains: new techniques for creating intelligent programs, using computers to understand the complexities of brain and mind, and, finally, contributing to philosophical debate.The techniques added to the AI repertoire are evolutionary programming, artificial life, and intelligent software agents. Evolutionary programming borrows from human genetic development in the sense that some variants of a program may have a better performance than others. It is possible to represent the design parameters of a system as a sequence of values resembling a chromosome. An evolutionary program tests a range of systems against a performance criterion (the fitness function). It chooses the chromosomes of various system pairs that have good fitness behaviour to combine them and create a new generation of systems. This gives rise to increasingly more able systems even to the extent that their design holds surprises for expert designers. Such mimicking of a major mechanism of biological life leads to the concept of artificial life. For example, UK entrepreneur Steve Grand includes in his 'Creatures' game simulations of some biochemical processes to produce societies of virtual (computer-bound but observable) creatures with realistic life cycles and social interactions. This allows the game player to take care of a virtual creature in a game that gets close to the problems of survival in real life. The more general study of intelligent software agents takes virtual creatures into domains where they could perform useful tasks such as finding desired data on the internet. They are little programs that store the needs of a user and trawl the World Wide Web for this desired information. Also a burgeoning interest is in societies of such agents to discover how cooperation between them may lead to the solution of problems in distributed domains. Translated to multiple interacting robots, agent studies lead to a better understanding of flocking behaviour and the way that this achieves goals for the flock.

A better understanding of the brain flows from the study of artificial neural networks (ANNs). Accepting that the brain is the most complex machine in existence, ANNs are now being used to isolate some of its structural features in order to begin to understand their interactions. For example, it has been possible to suggest a theoretical basis for understanding dyslexia, visual hallucinations under the influence of drugs, and the nature of visual awareness in general. The latter and grander ambition feeds a philosophical debate on whether machines could think like humans that has paralleled AI for its entire existence. The question was first raised by British mathematician Alan Turing in 1950. His celebrated test was based on the external behaviour of an AI machine and its ability to fool a human interlocutor into thinking that it too was human. This debate has now moved on to discuss whether a machine could ever be conscious. The main arguments against this come from a belief that consciousness, being a 'first-person' phenomenon, cannot be approached from the 'third-person' position which is inherent in all man-made designs. The contrary arguments are put by those who feel that by simulating with great care the function and structure of the brain it will be possible both to understand the mechanisms of consciousness and to transfer them to a machine.

Bibliography
• Aleksander, I. (2001). How to Build a Mind: Machines with Imagination.
• — —  and Morton, B. H. (1993). Neurons and Symbols.
• Boden, M. A. (ed.) (1996). Artificial Intelligence (2nd edn.).
• Crick, F. (1994). The Astonishing Hypothesis.
• Grand, S. (2000). Creation: Life and How to Make it.
• Searle, J. R. (1980). 'Minds brains and programs'. Behavioural and Brain Sciences, 3.
• Shannon, C. E. (1950). 'Programming a computer for playing chess'. Phil. Mag. 4.
• Tecuci, G. (1998). Building Intelligent Agents.
• Turing, A. M. (1950). 'Computing machinery and intelligence'. Mind, 59.

Margaret Boden

The science of making machines do the sorts of things that are done by human minds. Such things include holding a conversation, answering questions sensibly on the basis of incomplete knowledge, assembling another machine from its components given the blueprint, learning how to do things better, playing chess, writing or translating stories, understanding analogies, neurotically repressing knowledge that is too threatening to admit consciously, learning to classify visual or auditory patterns, composing a poem or a sonata, and recognizing the various things seen in a room — even an untidy and ill-lit room. AI helps one to realize how enormous is the background knowledge and thinking (computational) power needed to do even these everyday things.

The 'machines' in question are typically digital computers, but AI is not the study of computers. Rather, it is the study of intelligence in thought and action. Computers are its tools, because its theories are expressed as computer programs which are tested by being run on a machine. Some AI programs are lists of symbolic rules (if this is the case then do that, else do another ...). Others specify 'brainlike' networks made of many simple, interconnected, computational units. These types of AI are called traditional (or classical) and connectionist, respectively. They have differing, and largely complementary, strengths and weaknesses.

Other theories of intelligence are expressed verbally, either as psychological theories of thinking and behaviour, or as philosophical arguments about the nature of knowledge and purpose and the relation of mind to body (the mind–body problem). Because it approaches the same subject matter in different ways, AI is relevant to psychology and the philosophy of mind.

Similarly, attempts to write programs that can interpret the two-dimensional image from a TV camera in terms of the three-dimensional objects in the real world (or which can recognize photographs or drawings as representations of solid objects) help make explicit the range and subtlety of knowledge and unconscious inference that underlie our introspectively 'simple' experiences of seeing. Much of this knowledge is tacit (and largely innate) knowledge about the ways in which, given the laws of optics, physical surfaces of various kinds can give rise to specific visual images on a retina (or camera). Highly complex computational processes are needed to infer the nature of the physical object (or of its surfaces), on the basis of the two-dimensional image.

If we think of an AI system as a picture of a part of the mind, we must realize that a functioning program is more like a film of the mind than a portrait of it. Programming one's hunches about how the mind works is helpful in two ways. First, it enables one to express richly structured psychological theories in a rigorous, and testable, fashion. Second, it forces one to suggest specific hypotheses about precisely how a psychological change can come about. Even if (as in connectionist systems: see below) one only provides a learning rule, rather than telling the AI system precisely what to learn, that rule has to be rigorously expressed; a different rule will lead to different performance.

In general, it is easier to model logical and mathematical reasoning (which people find difficult) than to simulate high-level perception or language understanding (which we do more or less effortlessly). Significant progress has been made, for instance, in recognizing keywords and grammatical structure, and AI programs can even come up with respectable, though juvenile, puns and jokes. But many sentences, and jokes, assume a large amount of world knowledge, including culture-specific knowledge about sport, fashion, politics, soap operas ... the list is literally endless. There is little or no likelihood than an actual AI system could use language as well as we can, because it is too difficult to provide, and to structure, the relevant knowledge (much of it is tacit, and very difficult to bring into consciousness). But this need not matter, if all we want is a psychological theory that explains how these human capacities are possible. Similarly, research in AI has shown that highly complex, and typically unconscious, computational processes are needed to infer the nature of physical objects from the image reaching the retina/camera.

Traditional philosophical puzzles connected with the mind–body problem can often be illuminated by AI, because modelling a psychological phenomenon on a computer is a way of showing that and how it is possible for that phenomenon to arise in a physical system. For instance, people often feel that only a spiritual being (as opposed to a bodily one) could have purposes and try to achieve them, and the problem then arises of how the spiritual being, or mind, can possibly tell the body what to do, so that the body's hand can try to achieve the mind's purpose of, say, picking a daisy. It is relevant to ask whether, and how, a program can enable a machine to show the characteristic features of purpose. Is its behaviour guided by its idea of a future state? Is that idea sometimes illusory or mistaken (so that the 'daisy' is made of plastic, or is really a buttercup)? Does it symbolize what it is doing in terms of goals and subgoals (so that the picking of the daisy may be subordinate to the goal of stocking the classroom nature table)? Does it use this representation to help plan its actions (so that the daisies on the path outside the sweetshop are picked, rather than those by the petrol station)? Does it vary its means–end activities so as to achieve its goal in different circumstances (so that buttercups will do for the nature table if all the daisies have died)? Does it learn how to do so better (so that daisies for a daisy-chain are picked with long stalks)? Does it judge which purposes are the more important, or easier to achieve, and behave accordingly (if necessary, abandoning the daisy picking when a swarm of bees appears with an equally strong interest in the daisies)? Questions like these, asked with specific examples of functioning AI systems in mind, cannot fail to clarify the concept of purpose. Likewise, philosophical problems about the nature and criteria of knowledge can be clarified by reference to programs that process and use knowledge, so that AI is relevant to epistemology.

AI is concerned with mental processing in general, not just with mathematics and logical deduction. It includes computer models of perception, thought, motivation, and emotion. Emotion, for instance, is not just a feeling: emotions are scheduling mechanisms that have evolved to enable finite creatures with many potentially conflicting motives to choose what to do, when. (No matter how hungry one is, one had better stop eating and run away if faced by a tiger.) So a complex animal is going to need some form of computational interrupt, and some way of 'stacking' and realerting those unfulfilled intentions that shouldn't, or needn't, be abandoned. In human language users, motivational–emotional processing includes deliberately thought–out plans and contingency plans, and anticipation of possible outcomes from the various actions being considered.

One important variety of AI is connectionism, or artificial neural networks. Very few connectionist systems are implemented in fundamentally connectionist hardware. Most are simulated (as virtual machines) in digital computers. That is, the program does not list a sequence of symbolic rules but simulates many interconnected 'neurons', each of which does only very simple things. Connectionism enables a type of learning wherein the 'weights' on individual units in the network are gradually altered until recognition errors are minimized. Unlike learning in classical AI, the unfamiliar pattern need not be specifically described to the system before it can be learnt; however, it must be describable in the 'vocabulary' used for the system's input. Connectionism allows that beliefs and perceptions may be grounded on partly inconsistent evidence, and that most concepts are not strictly defined in terms of necessary and sufficient conditions. Many connectionist systems represent a concept as a pattern of activity across the whole network; the units eventually settle into a state of maximum, though not necessarily perfect, equilibrium. Connectionism is a powerful way of implementing pattern recognition and the 'intuitive' association of ideas. But it is very limited for implementing hierarchical structure of sequential processes, such as are involved in deliberate planning. Some AI research aims to develop 'hybrid' systems combining the strengths of traditional and connectionist AI. Certainly, the full range of adult human psychology cannot be captured by either of these approaches alone.

The main areas of AI include natural language understanding (see speech recognition by machine), machine vision (see pattern recognition), problem solving and game playing (see computer chess), robotics, automatic programming, and the development of programming languages. Among the practical applications most recently developed or currently being developed are medical diagnosis and treatment (where a program with specialist knowledge of, say, bacterial infections answers the questions of and elicits further relevant information from a general practitioner who is uncertain which drug to prescribe in a given case); prediction of share prices on the stock exchange; assessment of creditworthiness; speech analysis and speech synthesis; the composition of music, including jazz improvisation; location of mineral deposits, such as gold or oil; continuous route planning for car drivers; programs for playing chess, bridge, or Go, etc.; teaching some subject such as geography, or electronics, to students with differing degrees of understanding of the material to be explored; the automatic assembly of factory-made items, where the parts may have to be inspected first for various types of flaw and where they need not be accurately positioned at a precise point in the assembly line, as is needed for the automation in widespread use today; and the design of complex systems, whether electrical circuits or living spaces or some other, taking into account factors that may interact with each other in complicated ways (so that a mere 'checklist' program would not be adequate to solve the design problem).

An area closely related to AI is artificial life (A-life). This is a form of mathematical biology. It uses computational concepts and models to study (co-)evolution and self-organization, both of which apply to life in general, and to explain specific aspects of living things — such as navigation in insects or flocking in birds. (The dinosaurs in Jurassic Park were computer generated using simple A-life algorithms.) One example of A-life is evolutionary robotics, where the robot's neural network 'brain' and/or sensorimotor anatomy is not designed by hand but evolved over thousands of generations. The programs make random changes in their own rules, and a fitness function is applied, either automatically or manually, to select the best from the resulting examples; these are then used to breed the next generation. Some A-life scientists, but not all, accept 'strong' A-life: the view that a virtual creature, defined by computer software, could be genuinely alive. And some believe that A-life could help us to find an agreed definition of what 'life' is. All the minds we know of are embodied in living things, and some people argue that only a living thing could have a mind, or be intelligent. If that is right, then success in AI cannot be achieved without success in A-life. (In both cases, however, 'success' might be interpreted either as merely showing mindlike/lifelike behaviour or as being genuinely intelligent/alive.)

The social implications of AI are various. As with all technologies, there are potential applications which may prove bad, good, or ambiguous in human terms. A competent medical diagnosis program could be very useful, whereas a competent military application would be horrific for those at the receiving end, and a complex data-handling system could be well or ill used in many ways by individuals or governments. Then there is the question of what general implication AI will be seen to have for the commonly held 'image of man'. If it is interpreted by the public as implying that people are 'nothing but clockwork, really', then the indirect effects on self-esteem and social relations could be destructive of many of our most deeply held values. But it could (and should) be interpreted in a radically different and less dehumanizing way, as showing how it is possible for material systems (which, according to the biologist, we are) to possess such characteristic features of human psychology as subjectivity, purpose, freedom, and choice. The central theoretical concept in AI is representation, and AI workers ask how a (programmed) system constructs, adapts, and uses its inner representations in interpreting and changing its world. On this view, a programmed computer may be thought of as a subjective system (subject to illusion and error much as we are) functioning by way of its idiosyncratic view of the world. By analogy, then, it is no longer scientifically disreputable, as it has been thought to be for so long, to describe people in these radically subjective terms also. AI can therefore counteract the dehumanizing influence of the natural sciences that has been part of the mechanization of our world picture since the scientific revolution of the 16th and 17th centuries.

Bibliography
• Boden, M. A. (1987). Artificial Intelligence and Natural Man (2nd rev. edn.).
• — —  (1990). The Creative Mind.
• — —  (ed.) (1990). The Philosophy of Artificial Intelligence.
• Clark, A. J. (1990). Associative Engines.
• Cope, D. (2001). Virtual Music.
• Feigenbaum, E. A., and Feldman, J. (eds.) (1963). Computers and Thought.
• Jullam, J. (ed.) (1995). Hybrid Problems, Hybrid Solutions.
• Levy, S. J. (1992). Artificial Life.
• McClelland, J. L., and Rumelhart, D. E. (eds.) (1986). Parallel Distributed Processing: Explorations in the Microstructure of Cognition, 2 vols.
• Marr, D. A. (1980). Vision.
• Minsky, M. L. (1985). The Society of Mind.
• Whitby, B. (1996). Reflections on Artificial Intelligence.
• Winograd, T. (1972). Understanding Natural Language.

Ron Chrisley

Researchers in artificial intelligence attempt to design and create artefacts which have, or at least appear to have, mental properties: not just intelligences, but also perception, action, emotion, creativity, and consciousness.

1. Recent developments
2. The relevance of AI to understanding the mind

1. Recent developments

Since the mid-1980s, there has been sustained development of the core ideas of artificial intelligence, e.g. representation, planning, reasoning, natural language processing, machine learning, and perception. In addition, various subfields have emerged, such as research into agents (autonomous, independent systems, whether in hardware or software), distributed or multi-agent systems, coping with uncertainty, affective computing/models of emotion, and ontologies (systems of representing various kinds of entities in the world — achievements that, while new advances, are conceptually and methodologically continuous with the field of artificial intelligence as envisaged at the time of its modern genesis: the Dartmouth conference of 1956.

However, a substantial and growing proportion of research into artificial intelligence, while often building on the foundations just mentioned, has shifted its emphasis. This change in emphasis, inasmuch as it constitutes a conceptual break with those foundations, promises to make substantial contributions to our understanding and concepts of mind. It remains to be seen whether these contributions will replace or (as may seem more likely) merely supplement those already provided by what might be termed the 'Dartmouth approach' and its direct successors.

The new developments, which have their roots in the cybernetics work of the 1940s and 1950s as much as, if not more than, they do in mainstream AI, can be divided into two broad areas: adaptive systems and embodied/situated approaches. This is not to say that they are exclusive; much promising work, such as the field of evolutionary robotics, combines elements of both areas.Adaptive systems The 1980s saw a rise in the popularity of both neural networks (sometimes also called connectionist models) and genetic algorithms. Neural networks are systems comprising thousands or more of (usually simulated) simple processing units; the computational result of the network is determined by the input and the connections between the units, which may vary their ability to pass a signal from one unit to the next. Nearly all of these networks are adaptive in that they can learn. Learning typically consists in finding a set of connections that will make the network give the right output for each input in a given training set.

Genetic algorithms produce systems that perform well on some tasks by emulating natural selection. An initial random population of systems (whose properties are determined by a few parameters) are ranked according to their performance on the task; only the best performers are retained (selection). A new population is created by mutating or combining the parameters of the winners (reproduction and variation). Then the cycle repeats.

Although the importance of learning had been acknowledged since the earliest days of AI, these two approaches, despite their differences, had a common effect of making adaptivity absolutely central to AI.

While machine learning assumed conceptual building blocks with which to build learned structures, neural networks allowed for subsymbolic learning: the acquisition of the conceptual 'blocks' themselves, in a way that cannot be understood in terms of logical inference, and that may involve a continuous change of parameters, rather than discrete steps of accepting or rejecting sentences as being true or false. By allowing systems to construct their own 'take' on the world, AI researchers were able to begin overcoming the obstacles that were thrown up when they attempted to put adult human conceptual structures into systems that were quite different from us.

Standard AI methodology for giving some problem-solving capability to a machine had at first been: think about how you would solve the problem, write down the steps of your solution in a computer language, give the program to the machine to run. This was refined and extended in several ways. For example, the knowledge engineering approach asks an expert about the important facts of the domain, translates these into sentences in a knowledge representation language, gives these sentences to the machine, and lets the machine perform various forms of reasoning by manipulating these sentences. But it remained the case that, in these extensions of the basic AI methodology, the machine was limited to using the programmer's or expert's way of representing the world. By using adaptive approaches like artificial evolution, AI systems are no longer limited to solutions that humans can conceptualize — in fact the evolved or learned solutions are often inscrutable. Our concepts and intuitions might not be of much use in getting a six-legged robot to walk; our introspection might even lead us astray concerning the workings of our own minds. For both reasons, genetic algorithms are an impressive addition to the AI methodological toolbox.

However, along with these advantages come limitations. There is a general consensus that the simple, incremental methods of the adaptive approaches, while giving relatively good results for tasks closely related to perception and action, cannot scale up to tasks that require sophisticated, abstract, and conceptual abilities. Give a system some symbols and some rules for combining them, and it can potentially produce an infinite number of well-formed symbol structures — a feature that parallels human competence. But a neural network that has learned to produce a set of complex structures will usually fail to generalize this into a systematic competence to construct an infinite number of novel combinations. Genetic algorithms have similar limitations to their 'scaling up'. But even if these obstacles are overcome, and systems with advanced forms of mentality are created by these means, the very fact that we shall not have imposed our own concepts on them may render their behaviour itself inexplicable. What we do not need is another mind we cannot understand! With respect to AI's goal of adding to our understanding of the mind, adaptive (especially evolved) systems may be as much a part of the problem as a part of the solution (see section 2). And technological AI is also hindered if the systems it produces cannot be understood well enough to be trusted for use in the real world.Embodied and situated systems Embodied and situated approaches to AI investigate the role that the body and its sensorimotor processes (as opposed to symbols or representations on their own) can and do play in intelligent behaviour. Intelligence is viewed as the capacity for real-time, situated activity, typically inseparable from and often fully interleaved with perception and action. Further, it is by having a body that a system is situated in the world, and can thus exploit its relations to things in the world in order to perform tasks that might previously have been thought to require the manipulation of internal representations or data structures. For an example of embodied intelligence, suppose a child sees something of interest in front of him, points to it, turns his head back to get his mother's attention, and then returns his gaze to the front. He does not need to have some internal representation that stores the eye, neck, torso, etc. positions necessary to gaze on the item of interest; the child's arm itself will indicate where the child should look; the child's exploitation of his own embodiment obviates the need for him to store and access a complex inner symbolic structure. For an example of situated problem solving, suppose another child is solving a jigsaw puzzle. The child does not need to look at each piece intently, forming an internal representation of its shape, and then when all pieces have been examined, close her eyes and solve the puzzle in her head! Rather, the child can manipulate the pieces themselves, making it possible for her to perceive whether two of them will fit together. If nature has sometimes used these alternatives to complex inner symbol processing, then AI can (perhaps must) as well.

These are a cluster of other AI approaches that, while properly distinct from embodiment and situatedness, are nevertheless their natural allies.

(i) Some researchers have found it useful to turn away from discontinuous, atemporal, logic-based formalisms and instead use the continuous mathematics of change offered by dynamical systems theory as a way to characterize and design intelligent systems.
(ii) Some researchers have claimed that AI should, whenever possible, build systems working in the real world, with, for example, real cameras receiving real light, instead of relying on ray-traced simulations of light; a real-world AI system might exploit aspects of a situation we are not aware of and which we therefore do not incorporate in our simulations.
(iii) Some insist that AI should concentrate on building complete working systems, with simple but functioning and interacting perceptual, reasoning, learning, action, etc. systems, rather than working on developed yet isolated competences, as has been the method in the past.

Architectures A change of emphasis common to both the more and less traditional varieties of AI is a move away from a search for specific algorithms and representations, and toward a search for the architectures that support various forms of mentality. An architecture specifies how the various components of a system, which may in fact be representations or algorithms, fit together and interact in order to yield a working system. Thus, an architecture-based approach can render irrelevant many debates over which algorithm or representational scheme is 'best'.

2. The relevance of AI to understanding the mind

Why do AI? Of course, there are technological reasons. But are there scientific reasons? Can AI illuminate our understanding of the mind? The acts involved in bringing natural intelligences into the world do not (usually!) confer any insight into the nature of intelligence; why should one think the acts involved in creating artificial intelligence would be any more enlightening?

For one thing, not all AI eschews design to the extent that the genetic algorithm approach (above) does; most approaches involve the designer understanding, in advance, at least roughly how the constructed system works. AI need not got so far as to say 'if you can't build it, you can't understand it', but building an intelligence might at least help.

It is sometimes argued in return that the kind of systems that AI is likely to produce will be so different from naturally intelligent systems (e.g. they are not alive) that

(i) they will not shed much light on natural intelligence and
(ii) they will not be able to reach the heights that natural intelligence does.

Surely, these people conclude, if one is interested in intelligence and the mind, one should instead do neuroscience, or at least psychology?

One can defend the AI methodology for understanding natural intelligence by appealing to the history of understanding flight. Attempts both to achieve artificial flight and to understand natural flight failed as long as scientists tried to reproduce too closely what they saw in nature. It wasn't until scientists looked at simple, synthetic systems (such as Bernoulli's aerofoil), which could be arbitrarily manipulated and studied, that the general aerodynamic principles that underlie both artificial and natural flight could be identified. So also it may be that it is only by creating and interacting with simple (but increasingly complex) artificial systems that we will be able to uncover the general principles that will allow us both to construct artificial intelligence and understand natural intelligence.

(Published 2004)

Bibliography
• Bishop, C. M. (1995). Neural Networks for Pattern Recognition.
• Brooks, R. (1991). 'Intelligence without representation'. Artificial Intelligence, 47.
• Chrisley, R. (ed.) (2000). Artificial Intelligence: Critical Concepts.
• Clark, A. (1997). Being There: Putting Brain, Body and World Together Again.
• Franklin, S. (1995). Artificial Minds.
• Goldberg, D. E. (1989). Genetic Algorithms in Search, Optimization, and Machine Learning.
• Nilsson, N. (1998). Artificial Intelligence: A New Synthesis.
• Sharples, M., Hogg, D., Hutchison, C., Torrance, S., and Young, D. (1998). Computers and Thought: A Practical Introduction to Artificial Intelligence.
• Sloman, A. (1997). 'What sort of architecture is required for a human-like agent?' In Wooldridge, M., and Rao, A. (eds.), Foundations of Rational Agency.
• Smith, B. C. (1991). 'The owl and the electric encyclopedia'. Artificial Intelligence.

---