Chrisley, R. (2002, in press) "Artificial intelligence". In Gregory, R. (ed.) The Oxford Companion to the Mind (second edition). artificial intelligence (AI) Researchers in artificial intelligence attempt to design and create artefacts which have, or at least appear to have, mental properties: not just intelligence, but also perception, action, emotion, creativity, and consciousness. 1. Recent developments 1.1 Adaptive systems 1.2 Embodied/situated systems 1.3 Architectures 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 sub-fields 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 which, 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 40s and 50s 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. 1.1 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 task 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 sub-symbolic 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 in 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 generalise 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 don't need is another mind we can't 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. 1.2 Embodied and situated systems Embodied and situated approaches to AI investigate the role that the body and its sensory-motor 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 (and perhaps must) as well. There are a cluster of other AI approaches that, while properly distinct from embodiment and situatedness, are nevertheless their natural allies: 1) 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 characterise and design intelligent systems; 2) Some researchers have claimed that AI should, whenever possible, build systems working in the real world, with, e.g., real cameras receiving real light, instead of relying on e.g. 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; 3) 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. 1.3 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 (section 1.2) does; most approaches involve the designer understanding, in advance, at least roughly how the constructed system works. AI need not go 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 1) they will not shed much light on natural intelligence and 2) they won't 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. RLC References: 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. (1989) 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 47. 2015 words Ron Chrisley is a Senior Lecturer in the School of Cognitive and Computing Sciences at the University of Sussex, currently a Leverhulme Research Fellow on secondment to the School of Computer Science, University of Birmingham.