ARTIFICIAL INTELLIGENCE

Artificial Intelligence (AI) tries to enable computers to do the things that minds can do. These things include seeing pathways, picking things up, learning categories from experience, and using emotions to schedule one's actions—which many animals can do, too. Thus, human intelligence is not the sole focus of AI. Even terrestrial psychology is not the sole focus, because some people use AI to explore the range of all possible minds.

There are four major AI methodologies: symbolic AI, connectionism, situated robotics, and evolutionary programming (Russell and Norvig 2003). AI artifacts are correspondingly varied. They include both programs (including neural networks) and robots, each of which may be either designed in detail or largely evolved. The field is closely related to artificial life (A-Life), which aimsPage 346 | Top of Articleto throw light on biology much as some AI aims to throw light on psychology.

AI researchers are inspired by two different intellectual motivations, and while some people have both, most favor one over the other. On the one hand, many AI researchers seek solutions to technological problems, not caring whether these resemble human (or animal) psychology. They often make use of ideas about how people do things. Programs designed to aid/replace human experts, for example, have been hugely influenced by knowledge engineering, in which programmers try to discover what, and how, human experts are thinking when they do the tasks being modeled. But if these technological AI workers can find a nonhuman method, or even a mere trick (a kludge) to increase the power of their program, they will gladly use it.

Technological AI has been hugely successful. It has entered administrative, financial, medical, and manufacturing practice at countless different points. It is largely invisible to the ordinary person, lying behind some deceptively simple human-computer interface or being hidden away inside a car or refrigerator. Many procedures taken for granted within current computer science were originated within AI (pattern-recognition and image-processing, for example).

On the other hand, AI researchers may have a scientific aim. They may want their programs or robots to help people understand how human (or animal) minds work. They may even ask how intelligence in general is possible, exploring the space of possible minds. The scientific approach—psychological AI—is the more relevant for philosophers (Boden 1990, Copeland 1993, Sloman 2002). It is also central to cognitive science, and to computationalism.

Considered as a whole, psychological AI has been less obviously successful than technological AI. This is partly because the tasks it tries to achieve are often more difficult. In addition, it is less clear—for philosophical as well as empirical reasons—what should be counted as success.

Symbolic AI is also known as classical AI and as GOFAI—short for John Haugeland's label "Good Old-Fashioned AI" (1985). It models mental processes as the step-by-step information processing of digital computers. Thinking is seen as symbol-manipulation, as (formal) computation over (formal) representations. Some GOFAI programs are explicitly hierarchical, consisting of procedures and subroutines specified at different levels. These define a hierarchically structured search-space, which may be astronomical in size. Rules of thumb, or heuristics, are typically provided to guide the search—by excluding certain areas of possibility, and leading the program to focus on others. The earliest AI programs were like this, but the later methodology of object-oriented programming is similar.

Certain symbolic programs, namely production systems, are implicitly hierarchical. These consist of sets of logically separate if-then (condition-action) rules, or productions, defining what actions should be taken in response to specific conditions. An action or condition may be unitary or complex, in the latter case being defined by a conjunction of several mini-actions or mini-conditions. And a production may function wholly within computer memory (to set a goal, for instance, or to record a partial parsing) or outside it (via input/output devices such as cameras or keyboards).

Another symbolic technique, widely used in natural language processing (NLP) programs, involves augmented transition networks, or ATNs. These avoid explicit backtracking by using guidance at each decision-point to decide which question to ask and/or which path to take.

GOFAI methodology is used for developing a wide variety of language-using programs and problem-solvers. The more precisely and explicitly a problem-domain can be defined, the more likely it is that a symbolic program can be used to good effect. Often, folk-psychological categories and/or specific propositions are explicitly represented in the system. This type of AI, and the forms of computational psychology based on it, is defended by the philosopher Jerry Fodor (1988).

GOFAI models (whether technological or scientific) include robots, planning programs, theorem-provers, learning programs, question-answerers, data-mining systems, machine translators, expert systems of many different kinds, chess players, semantic networks, and analogy machines. In addition, a host of software agents—specialist mini-programs that can aid a human being to solve a problem—are implemented in this way. And an increasingly important area of research is distributed AI, in which cooperation occurs between many relatively simple individuals—which may be GOFAI agents (or neural-network units, or situated robots).

The symbolic approach is used also in modeling creativity in various domains (Boden 2004, Holland et al. 1986). These include musical composition and expressive performance, analogical thinking, line-drawing, painting,Page 347 | Top of Articlearchitectural design, storytelling (rhetoric as well as plot), mathematics, and scientific discovery. In general, the relevant aesthetic/theoretical style must be specified clearly, so as to define a space of possibilities that can be fruitfully explored by the computer. To what extent the exploratory procedures can plausibly be seen as similar to those used by people varies from case to case.

Connectionist systems, which became widely visible in the mid-1980s, are different. They compute not by following step-by-step programs but by using large numbers of locally connected (associative) computational units, each one of which is simple. The processing is bottom-up rather than top-down.

Connectionism is sometimes said to be opposed to AI, although it has been part of AI since its beginnings in the 1940s (McCulloch and Pitts 1943, Pitts and McCulloch 1947). What connectionism is opposed to, rather, is symbolic AI. Yet even here, opposed is not quite the right word, since hybrid systems exist that combine both methodologies. Moreover, GOFAI devotees such as Fodor see connectionism as compatible with GOFAI, claiming that it concerns how symbolic computation can be implemented (Fodor and Pylyshyn 1988).

Two largely separate AI communities began to emerge in the late 1950s (Boden forthcoming). The symbolic school focused on logic and Turing-computation, whereas the connectionist school focused on associative, and often probabilistic, neural networks. (Most connectionist systems are connectionist virtual machines, implemented in von Neumann computers; only a few are built in dedicated connectionist hardware.) Many people remained sympathetic to both schools. But the two methodologies are so different in practice that most hands-on AI researchers use either one or the other.

There are different types of connectionist systems. Most philosophical interest, however, has focused on networks that do parallel distributed processing, or PDP (Clark 1989, Rumelhart and McClelland 1986). In essence, PDP systems are pattern recognizers. Unlike brittle GOFAI programs, which often produce nonsense if provided with incomplete or part-contradictory information, they show graceful degradation. That is, the input patterns can be recognized (up to a point) even if they are imperfect.

A PDP network is made up of subsymbolic units, whose semantic significance cannot easily be expressed in terms of familiar semantic content, still less propositions. (Some GOFAI programs employ subsymbolic units, but most do not.) That is, no single unit codes for a recognizable concept, such as dog or cat. These concepts are represented, rather, by the pattern of activity distributed over the entire network.

Because the representation is not stored in a single unit but is distributed over the whole network, PDP systems can tolerate imperfect data. (Some GOFAI systems can do so too, but only if the imperfections are specifically foreseen and provided for by the programmer.) Moreover, a single subsymbolic unit may mean one thing in one input-context and another in another. What the network as a whole can represent depends on what significance the designer has decided to assign to the input-units. For instance, some input-units are sensitive to light (or to coded information about light), others to sound, others to triads of phonological categories … and so on.

Most PDP systems can learn. In such cases, the weights on the links of PDP units in the hidden layer (between the input-layer and the output-layer) can be altered by experience, so that the network can learn a pattern merely by being shown many examples of it. (A GOFAI learning-program, in effect, has to be told what to look for beforehand, and how.) Broadly, the weight on an excitatory link is increased by every coactivation of the two units concerned: cells that fire together, wire together.

These two AI approaches have complementary strengths and weaknesses. For instance, symbolic AI is better at modeling hierarchy and strong constraints, whereas connectionism copes better with pattern recognition, especially if many conflicting—and perhaps incomplete—constraints are relevant. Despite having fervent philosophical champions on both sides, neither methodology is adequate for all of the tasks dealt with by AI scientists. Indeed, much research in connectionism has aimed to restore the lost logical strengths of GOFAI to neural networks—with only limited success by the beginning of the twenty-first century.

Another, and more recently popular, AI methodology is situated robotics (Brooks 1991). Like connectionism, this was first explored in the 1950s. Situated robots are described by their designers as autonomous systems embedded in their environment (Heidegger is sometimes cited). Instead of planning their actions, as classical robots do, situated robots react directly to environmental cues. One might say that they are embodied production systems, whose if-then rules are engineered rather than programmed, and whose conditions lie in the externalPage 348 | Top of Articleenvironment, not inside computer memory. Although—unlike GOFAI robots—they contain no objective representations of the world, some of them do construct temporary, subject-centered (deictic) representations.

The main aim of situated roboticists in the mid-1980s, such as Rodney Brooks, was to solve/avoid the frame problem that had bedeviled GOFAI (Pylyshyn 1987). GOFAI planners and robots had to anticipate all possible contingencies, including the side effects of actions taken by the system itself, if they were not to be defeated by unexpected—perhaps seemingly irrelevant—events. This was one of the reasons given by Hubert Dreyfus (1992) in arguing that GOFAI could not possibly succeed: Intelligence, he said, is unformalizable. Several ways of implementing nonmonotonic logics in GOFAI were suggested, allowing a conclusion previously drawn by faultless reasoning to be negated by new evidence. But because the general nature of that new evidence had to be foreseen, the frame problem persisted.

Brooks argued that reasoning shouldn't be employed at all: the system should simply react appropriately, in a reflex fashion, to specific environmental cues. This, he said, is what insects do—and they are highly successful creatures. (Soon, situated robotics was being used, for instance, to model the six-legged movement of cockroaches.) Some people joked that AI stood for artificialinsects, not artificial intelligence. But the joke carried a sting: Many argued that much human thinking needs objective representations, so the scope for situated robotics was strictly limited.

In evolutionary programming, genetic algorithms (GAs) are used by a program to make random variations in its own rules. The initial rules, before evolution begins, either do not achieve the task in question or do so only inefficiently; sometimes, they are even chosen at random.

The variations allowed are broadly modeled on biological mutations and crossovers, although more unnatural types are sometimes employed. The most successful rules are automatically selected, and then varied again. This is more easily said than done: The breakthrough in GA methodology occurred when John Holland (1992) defined an automatic procedure for recognizing which rules, out of a large and simultaneously active set, were those most responsible for whatever level of success the evolving system had just achieved.

Selection is done by some specific fitness criterion, predefined in light of the task the programmer has in mind. Unlike GOFAI systems, a GA program contains no explicit representation of what it is required to do: its task is implicit in the fitness criterion. (Similarly, living things have evolved to do what they do without knowing what that is.) After many generations, the GA system may be well-adapted to its task. For certain types of tasks, it can even find the optimal solution.

This AI method is used to develop both symbolic and connectionist AI systems. And it is applied both to abstract problem-solving (mathematical optimization, for instance, or the synthesis of new pharmaceutical molecules) and to evolutionary robotics—wherein the brain and/or sensorimotor anatomy of robots evolve within a specific task-environment.

It is also used for artistic purposes, in the composition of music or the generation of new visual forms. In these cases, evolution is usually interactive. That is, the variation is done automatically but the selection is done by a human being—who does not need to (and usually could not) define, or even name, the aesthetic fitness criteria being applied.

AI is a close cousin of A-Life (Boden 1996). This is a form of mathematical biology, which employs computer simulation and situated robotics to study the emergence of complexity in self-organizing, self-reproducing, adaptive systems. (A caveat: much as some AI is purely technological in aim, so is some A-Life; the research of most interest to philosophers is the scientifically oriented type.)

The key concepts of A-Life date back to the early 1950s. They originated in theoretical work on self-organizing systems of various kinds, including diffusion equations and cellular automata (by Alan Turing and John von Neumann respectively), and in early self-equilibrating machines and situated robots (built by W. Ross Ashby and W. Grey Walter). But A-Life did not flourish until the late 1980s, when computing power at last sufficed to explore these theoretical ideas in practice.

Much A-Life work focuses on specific biological phenomena, such as flocking, cooperation in ant colonies, or morphogenesis—from cell-differentiation to the formation of leopard spots or tiger stripes. But A-Life also studies general principles of self-organization in biology: evolution and coevolution, reproduction, and metabolism. In addition, it explores the nature of life as such—life as it could be, not merely life as it is.

A-Life workers do not all use the same methodology, but they do eschew the top-down methods of GOFAI. SituatedPage 349 | Top of Articleand evolutionary robotics, and GA-generated neural networks, too, are prominent approaches within the field. But not all A-Life systems are evolutionary. Some demonstrate how a small number of fixed, and simple, rules can lead to self-organization of an apparently complex kind.

Many A-Lifers take pains to distance themselves from AI. But besides their close historical connections, AI and A-Life are philosophically related in virtue of the linkage between life and mind. It is known that psychological properties arise in living things, and some people argue (or assume) that they can arise only in living things. Accordingly, the whole of AI could be regarded as a subarea of A-Life. Indeed, some people argue that success in AI (even in technological AI) must await, and build on, success in A-Life.

Whichever of the two AI motivations—technological or psychological—is in question, the name of the field is misleading in three ways. First, the term intelligence is normally understood to cover only a subset of what AI workers are trying to do. Second, intelligence is often supposed to be distinct from emotion, so that AI is assumed to exclude work on that. And third, the name implies that a successful AI system would really be intelligent—a philosophically controversial claim that AIresearchers do not have to endorse (though some do).

As for the first point, people do not normally regard vision or locomotion as examples of intelligence. Many people would say that speaking one's native language is not a case of intelligence either, except in comparison with nonhuman species; and common sense is sometimes contrasted with intelligence. The term is usually reserved for special cases of human thought that show exceptional creativity and subtlety, or which require many years of formal education. Medical diagnosis, scientific or legal reasoning, playing chess, and translating from one language to another are typically regarded as difficult, thus requiring intelligence. And these tasks were the main focus of research when AI began. Vision, for example, was assumed to be relatively straightforward—not least, because many nonhuman animals have it too. It gradually became clear, however, that everyday capacities such as vision and locomotion are vastly more complex than had been supposed. The early definition of AI as programming computers to do things that involve intelligence when done by people was recognized as misleading, and eventually dropped.

Similarly, intelligence is often opposed to emotion. Many people assume that AI could never model that. However, crude examples of such models existed in the early 1960s, and emotion was recognized by a high priest of AI, Herbert Simon, as being essential to any complex intelligence. Later, research in the computational philosophy (and modeling) of affect showed that emotions have evolved as scheduling mechanisms for systems with many different, and potentially conflicting, purposes (Minsky 1985, and Web site). When AI began, it was difficult enough to get a program to follow one goal (with its subgoals) intelligently—any more than that was essentially impossible. For this reason, among others, AI modeling of emotion was put on the back burner for about thirty years. By the 1990s, however, it had become a popular focus of AI research, and of neuroscience and philosophy too.

The third point raises the difficult question—which many AI practitioners leave open, or even ignore—of whether intentionality can properly be ascribed to any conceivable program/robot (Newell 1980, Dennett 1987, Harnad 1991).

Could some NLP programs really understand the sentences they parse and the words they translate? Or can a visuo-motor circuit evolved within a robot's neural-network brain truly be said to represent the environmental feature to which it responds? If a program, in practice, could pass the Turing Test, could it truly be said to think? More generally, does it even make sense to say that AImay one day achieve artificially produced (but nonetheless genuine) intelligence?

For the many people in the field who adopt some form of functionalism, the answer in each case is: In principle, yes. This applies for those who favor the physical symbol system hypothesis or intentional systems theory. Others adopt connectionist analyses of concepts, and of their development from nonconceptual content. Functionalism is criticized by many writers expert in neuroscience, who claim that its core thesis of multiple realizability is mistaken. Others criticize it at an even deeper level: a growing minority (especially in A-Life) reject neo-Cartesian approaches in favor of philosophies of embodiment, such as phenomenology or autopoiesis.

Part of the reason why such questions are so difficult is that philosophers disagree about what intentionality is, even in the human case. Practitioners of psychological AI generally believe that semantic content, or intentionality,Page 350 | Top of Articlecan be naturalized. But they differ about how this can be done.

For instance, a few practitioners of AI regard computation and intentionality as metaphysically inseparable (Smith 1996). Others ascribe meaning only to computations with certain causal consequences and provenance, or grounding. John Searle argues that AI cannot capture intentionality, because—at base—it is concerned with the formal manipulation of formal symbols. And for those who accept some form of evolutionary semantics, only evolutionary robots could embody meaning (Searle, 1980).