6 Queztion In Artificial Inteligence

CHAPTER 1 / AI: HISTORY AND APPLICATIONS 7

Pascal’s successes with calculating machines inspired Gottfried Wilhelm von Leibniz in 1694 to complete a working machine that become known as the Leibniz Wheel. It inte- grated a moveable carriage and hand crank to drive wheels and cylinders that performed the more complex operations of multiplication and division. Leibniz was also fascinated by the possibility of a automated logic for proofs of propositions. Returning to Bacon’s entity specification algorithm, where concepts were characterized as the collection of their necessary and sufficient features, Liebniz conjectured a machine that could calculate with these features to produce logically correct conclusions. Liebniz (1887) also envisioned a machine, reflecting modern ideas of deductive inference and proof, by which the produc- tion of scientific knowledge could become automated, a calculus for reasoning.

The seventeenth and eighteenth centuries also saw a great deal of discussion of episte- mological issues; perhaps the most influential was the work of René Descartes, a central figure in the development of the modern concepts of thought and theories of mind. In his Meditations, Descartes (1680) attempted to find a basis for reality purely through intro- spection. Systematically rejecting the input of his senses as untrustworthy, Descartes was forced to doubt even the existence of the physical world and was left with only the reality of thought; even his own existence had to be justified in terms of thought: “Cogito ergo sum” (I think, therefore I am). After he established his own existence purely as a thinking entity, Descartes inferred the existence of God as an essential creator and ultimately reas- serted the reality of the physical universe as the necessary creation of a benign God.

We can make two observations here: first, the schism between the mind and the phys- ical world had become so complete that the process of thinking could be discussed in iso- lation from any specific sensory input or worldly subject matter; second, the connection between mind and the physical world was so tenuous that it required the intervention of a benign God to support reliable knowledge of the physical world! This view of the duality between the mind and the physical world underlies all of Descartes’s thought, including his development of analytic geometry. How else could he have unified such a seemingly worldly branch of mathematics as geometry with such an abstract mathematical frame- work as algebra?

Why have we included this mind/body discussion in a book on artificial intelligence? There are two consequences of this analysis essential to the AI enterprise:

1. By attempting to separate the mind from the physical world, Descartes and related thinkers established that the structure of ideas about the world was not necessar- ily the same as the structure of their subject matter. This underlies the methodol- ogy of AI, along with the fields of epistemology, psychology, much of higher mathematics, and most of modern literature: mental processes have an existence of their own, obey their own laws, and can be studied in and of themselves.

2. Once the mind and the body are separated, philosophers found it necessary to find a way to reconnect the two, because interaction between Descartes mental, res cogitans, and physical, res extensa, is essential for human existence.

Although millions of words have been written on this mind–body problem, and numerous solutions proposed, no one has successfully explained the obvious interactions between mental states and physical actions while affirming a fundamental difference

 

 

8 PART I / ARTIFICIAL INTELLIGENCE: ITS ROOTS AND SCOPE

between them. The most widely accepted response to this problem, and the one that provides an essential foundation for the study of AI, holds that the mind and the body are not fundamentally different entities at all. On this view, mental processes are indeed achieved by physical systems such as brains (or computers). Mental processes, like physi- cal processes, can ultimately be characterized through formal mathematics. Or, as acknowledged in his Leviathan by the 17th century English philosopher Thomas Hobbes (1651), “By ratiocination, I mean computation”.

1.1.2 AI and the Rationalist and Empiricist Traditions

Modern research issues in artificial intelligence, as in other scientific disciplines, are formed and evolve through a combination of historical, social, and cultural pressures. Two of the most prominent pressures for the evolution of AI are the empiricist and rationalist traditions in philosophy.

The rationalist tradition, as seen in the previous section, had an early proponent in Plato, and was continued on through the writings of Pascal, Descartes, and Liebniz. For the rationalist, the external world is reconstructed through the clear and distinct ideas of a mathematics. A criticism of this dualistic approach is the forced disengagement of repre- sentational systems from their field of reference. The issue is whether the meaning attrib- uted to a representation can be defined independent of its application conditions. If the world is different from our beliefs about the world, can our created concepts and symbols still have meaning?

Many AI programs have very much of this rationalist flavor. Early robot planners, for example, would describe their application domain or “world” as sets of predicate calculus statements and then a “plan” for action would be created through proving theorems about this “world” (Fikes et al. 1972, see also Section 8.4). Newell and Simon’s Physical Symbol System Hypothesis (Introduction to Part II and Chapter 16) is seen by many as the arche- type of this approach in modern AI. Several critics have commented on this rationalist bias as part of the failure of AI at solving complex tasks such as understanding human lan- guages (Searle 1980, Winograd and Flores 1986, Brooks 1991a).

Rather than affirming as “real” the world of clear and distinct ideas, empiricists con- tinue to remind us that “nothing enters the mind except through the senses”. This con- straint leads to further questions of how the human can possibly perceive general concepts or the pure forms of Plato’s cave (Plato 1961). Aristotle was an early empiricist, emphasiz- ing in his De Anima, the limitations of the human perceptual system. More modern empir- icists, especially Hobbes, Locke, and Hume, emphasize that knowledge must be explained through an introspective but empirical psychology. They distinguish two types of mental phenomena perceptions on one hand and thought, memory, and imagination on the other. The Scots philosopher, David Hume, for example, distinguishes between impressions and ideas. Impressions are lively and vivid, reflecting the presence and existence of an exter- nal object and not subject to voluntary control, the qualia of Dennett (2005). Ideas on the other hand, are less vivid and detailed and more subject to the subject’s voluntary control.

Given this distinction between impressions and ideas, how can knowledge arise? For Hobbes, Locke, and Hume the fundamental explanatory mechanism is association.

 

 

CHAPTER 1 / AI: HISTORY AND APPLICATIONS 9

Particular perceptual properties are associated through repeated experience. This repeated association creates a disposition in the mind to associate the corresponding ideas, a pre- curser of the behaviorist approach of the twentieth century. A fundamental property of this account is presented with Hume’s skepticism. Hume’s purely descriptive account of the origins of ideas cannot, he claims, support belief in causality. Even the use of logic and induction cannot be rationally supported in this radical empiricist epistemology.

In An Inquiry Concerning Human Understanding (1748), Hume’s skepticism extended to the analysis of miracles. Although Hume didn’t address the nature of miracles directly, he did question the testimony-based belief in the miraculous. This skepticism, of course, was seen as a direct threat by believers in the bible as well as many other purvey- ors of religious traditions. The Reverend Thomas Bayes was both a mathematician and a minister. One of his papers, called Essay towards Solving a Problem in the Doctrine of Chances (1763) addressed Hume’s questions mathematically. Bayes’ theorem demon- strates formally how, through learning the correlations of the effects of actions, we can determine the probability of their causes.

The associational account of knowledge plays a significant role in the development of AI representational structures and programs, for example, in memory organization with semantic networks and MOPS and work in natural language understanding (see Sections 7.0, 7.1, and Chapter 15). Associational accounts have important influences of machine learning, especially with connectionist networks (see Section 10.6, 10.7, and Chapter 11). Associationism also plays an important role in cognitive psychology including the sche- mas of Bartlett and Piaget as well as the entire thrust of the behaviorist tradition (Luger 1994). Finally, with AI tools for stochastic analysis, including the Bayesian belief network (BBN) and its current extensions to first-order Turing-complete systems for stochastic modeling, associational theories have found a sound mathematical basis and mature expressive power. Bayesian tools are important for research including diagnostics, machine learning, and natural language understanding (see Chapters 5 and 13).

Immanuel Kant, a German philosopher trained in the rationalist tradition, was strongly influenced by the writing of Hume. As a result, he began the modern synthesis of these two traditions. Knowledge for Kant contains two collaborating energies, an a priori component coming from the subject’s reason along with an a posteriori component com- ing from active experience. Experience is meaningful only through the contribution of the subject. Without an active organizing form proposed by the subject, the world would be nothing more than passing transitory sensations. Finally, at the level of judgement, Kant claims, passing images or representations are bound together by the active subject and taken as the diverse appearances of an identity, of an “object”. Kant’s realism began the modern enterprise of psychologists such as Bartlett, Brunner, and Piaget. Kant’s work influences the modern AI enterprise of machine learning (Section IV) as well as the con- tinuing development of a constructivist epistemology (see Chapter 16).

1.1.3 The Development of Formal Logic

Once thinking had come to be regarded as a form of computation, its formalization and eventual mechanization were obvious next steps. As noted in Section 1.1.1,

 

 

10 PART I / ARTIFICIAL INTELLIGENCE: ITS ROOTS AND SCOPE

Gottfried Wilhelm von Leibniz, with his Calculus Philosophicus, introduced the first sys- tem of formal logic as well as proposed a machine for automating its tasks (Leibniz 1887). Furthermore, the steps and stages of this mechanical solution can be represented as move- ment through the states of a tree or graph. Leonhard Euler, in the eighteenth century, with his analysis of the “connectedness” of the bridges joining the riverbanks and islands of the city of Königsberg (see the introduction to Chapter 3), introduced the study of representa- tions that can abstractly capture the structure of relationships in the world as well as the discrete steps within a computation about these relationships (Euler 1735).

The formalization of graph theory also afforded the possibility of state space search, a major conceptual tool of artificial intelligence. We can use graphs to model the deeper structure of a problem. The nodes of a state space graph represent possible stages of a problem solution; the arcs of the graph represent inferences, moves in a game, or other steps in a problem solution. Solving the problem is a process of searching the state space graph for a path to a solution (Introduction to II and Chapter 3). By describing the entire space of problem solutions, state space graphs provide a powerful tool for measuring the structure and complexity of problems and analyzing the efficiency, correctness, and gener- ality of solution strategies.

As one of the originators of the science of operations research, as well as the designer of the first programmable mechanical computing machines, Charles Babbage, a nine- teenth century mathematician, may also be considered an early practitioner of artificial intelligence (Morrison and Morrison 1961). Babbage’s difference engine was a special- purpose machine for computing the values of certain polynomial functions and was the forerunner of his analytical engine. The analytical engine, designed but not successfully constructed during his lifetime, was a general-purpose programmable computing machine that presaged many of the architectural assumptions underlying the modern computer.

In describing the analytical engine, Ada Lovelace (1961), Babbage’s friend, sup- porter, and collaborator, said:

We may say most aptly that the Analytical Engine weaves algebraical patterns just as the Jac- quard loom weaves flowers and leaves. Here, it seems to us, resides much more of originality than the difference engine can be fairly entitled to claim.

Babbage’s inspiration was his desire to apply the technology of his day to liberate humans from the drudgery of making arithmetic calculations. In this sentiment, as well as with his conception of computers as mechanical devices, Babbage was thinking in purely nineteenth century terms. His analytical engine, however, also included many modern notions, such as the separation of memory and processor, the store and the mill in Bab- bage’s terms, the concept of a digital rather than analog machine, and programmability based on the execution of a series of operations encoded on punched pasteboard cards. The most striking feature of Ada Lovelace’s description, and of Babbage’s work in gen- eral, is its treatment of the “patterns” of algebraic relationships as entities that may be studied, characterized, and finally implemented and manipulated mechanically without concern for the particular values that are finally passed through the mill of the calculating machine. This is an example implementation of the “abstraction and manipulation of form” first described by Aristotle and Liebniz.

 

 

CHAPTER 1 / AI: HISTORY AND APPLICATIONS 11

The goal of creating a formal language for thought also appears in the work of George Boole, another nineteenth-century mathematician whose work must be included in any discussion of the roots of artificial intelligence (Boole 1847, 1854). Although he made contributions to a number of areas of mathematics, his best known work was in the mathematical formalization of the laws of logic, an accomplishment that forms the very heart of modern computer science. Though the role of Boolean algebra in the design of logic circuitry is well known, Boole’s own goals in developing his system seem closer to those of contemporary AI researchers. In the first chapter of An Investigation of the Laws of Thought, on which are founded the Mathematical Theories of Logic and Probabilities, Boole (1854) described his goals as

to investigate the fundamental laws of those operations of the mind by which reasoning is performed: to give expression to them in the symbolical language of a Calculus, and upon this foundation to establish the science of logic and instruct its method; …and finally to collect from the various elements of truth brought to view in the course of these inquiries some proba- ble intimations concerning the nature and constitution of the human mind.

The importance of Boole’s accomplishment is in the extraordinary power and sim- plicity of the system he devised: three operations, “AND” (denoted by

 

∗ or

 

∧), “OR” (denoted by

 

+ or

 

∨), and “NOT” (denoted by

 

¬), formed the heart of his logical calculus. These operations have remained the basis for all subsequent developments in formal logic, including the design of modern computers. While keeping the meaning of these symbols nearly identical to the corresponding algebraic operations, Boole noted that “the Symbols of logic are further subject to a special law, to which the symbols of quantity, as such, are not subject”. This law states that for any X, an element in the algebra, X

 

∗X

 

=X (or that once something is known to be true, repetition cannot augment that knowledge). This led to the characteristic restriction of Boolean values to the only two numbers that may satisfy this equation: 1 and 0. The standard definitions of Boolean multiplication (AND) and addition (OR) follow from this insight.

Boole’s system not only provided the basis of binary arithmetic but also demonstrated that an extremely simple formal system was adequate to capture the full power of logic. This assumption and the system Boole developed to demonstrate it form the basis of all modern efforts to formalize logic, from Russell and Whitehead’s Principia Mathematica (Whitehead and Russell 1950), through the work of Turing and Gödel, up to modern auto- mated reasoning systems.

Gottlob Frege, in his Foundations of Arithmetic (Frege 1879, 1884), created a mathematical specification language for describing the basis of arithmetic in a clear and precise fashion. With this language Frege formalized many of the issues first addressed by Aristotle’s Logic. Frege’s language, now called the first-order predicate calculus, offers a tool for describing the propositions and truth value assignments that make up the elements of mathematical reasoning and describes the axiomatic basis of “meaning” for these expressions. The formal system of the predicate calculus, which includes predicate sym- bols, a theory of functions, and quantified variables, was intended to be a language for describing mathematics and its philosophical foundations. It also plays a fundamental role in creating a theory of representation for artificial intelligence (Chapter 2). The first-order

 

 

12 PART I / ARTIFICIAL INTELLIGENCE: ITS ROOTS AND SCOPE

predicate calculus offers the tools necessary for automating reasoning: a language for expressions, a theory for assumptions related to the meaning of expressions, and a logi- cally sound calculus for inferring new true expressions.

Whitehead and Russell’s (1950) work is particularly important to the foundations of AI, in that their stated goal was to derive the whole of mathematics through formal opera- tions on a collection of axioms. Although many mathematical systems have been con- structed from basic axioms, what is interesting is Russell and Whitehead’s commitment to mathematics as a purely formal system. This meant that axioms and theorems would be treated solely as strings of characters: proofs would proceed solely through the application of well-defined rules for manipulating these strings. There would be no reliance on intu- ition or the meaning of theorems as a basis for proofs. Every step of a proof followed from the strict application of formal (syntactic) rules to either axioms or previously proven the- orems, even where traditional proofs might regard such a step as “obvious”. What “mean- ing” the theorems and axioms of the system might have in relation to the world would be independent of their logical derivations. This treatment of mathematical reasoning in purely formal (and hence mechanical) terms provided an essential basis for its automation on physical computers. The logical syntax and formal rules of inference developed by Russell and Whitehead are still a basis for automatic theorem-proving systems, presented in Chapter 14, as well as for the theoretical foundations of artificial intelligence.

Alfred Tarski is another mathematician whose work is essential to the foundations of AI. Tarski created a theory of reference wherein the well-formed formulae of Frege or Russell and Whitehead can be said to refer, in a precise fashion, to the physical world (Tarski 1944, 1956; see Chapter 2). This insight underlies most theories of formal seman- tics. In his paper The Semantic Conception of Truth and the Foundation of Semantics, Tar- ski describes his theory of reference and truth value relationships. Modern computer scientists, especially Scott, Strachey, Burstall (Burstall and Darlington 1977), and Plotkin have related this theory to programming languages and other specifications for computing.

Although in the eighteenth, nineteenth, and early twentieth centuries the formaliza- tion of science and mathematics created the intellectual prerequisite for the study of artifi- cial intelligence, it was not until the twentieth century and the introduction of the digital computer that AI became a viable scientific discipline. By the end of the 1940s electronic digital computers had demonstrated their potential to provide the memory and processing power required by intelligent programs. It was now possible to implement formal reason- ing systems on a computer and empirically test their sufficiency for exhibiting intelli- gence. An essential component of the science of artificial intelligence is this commitment to digital computers as the vehicle of choice for creating and testing theories of intelligence.

Digital computers are not merely a vehicle for testing theories of intelligence. Their architecture also suggests a specific paradigm for such theories: intelligence is a form of information processing. The notion of search as a problem-solving methodology, for example, owes more to the sequential nature of computer operation than it does to any biological model of intelligence. Most AI programs represent knowledge in some formal language that is then manipulated by algorithms, honoring the separation of data and program fundamental to the von Neumann style of computing. Formal logic has emerged as an important representational tool for AI research, just as graph theory plays an indis-

 

 

CHAPTER 1 / AI: HISTORY AND APPLICATIONS 13

pensable role in the analysis of problem spaces as well as providing a basis for semantic networks and similar models of semantic meaning. These techniques and formalisms are discussed in detail throughout the body of this text; we mention them here to emphasize the symbiotic relationship between the digital computer and the theoretical underpinnings of artificial intelligence.

We often forget that the tools we create for our own purposes tend to shape our conception of the world through their structure and limitations. Although seemingly restrictive, this interaction is an essential aspect of the evolution of human knowledge: a tool (and scientific theories are ultimately only tools) is developed to solve a particular problem. As it is used and refined, the tool itself seems to suggest other applications, leading to new questions and, ultimately, the development of new tools.

1.1.4 The Turing Test

One of the earliest papers to address the question of machine intelligence specifically in relation to the modern digital computer was written in 1950 by the British mathematician Alan Turing. Computing Machinery and Intelligence (Turing 1950) remains timely in both its assessment of the arguments against the possibility of creating an intelligent computing machine and its answers to those arguments. Turing, known mainly for his contributions to the theory of computability, considered the question of whether or not a machine could actually be made to think. Noting that the fundamental ambiguities in the question itself (what is thinking? what is a machine?) precluded any rational answer, he proposed that the question of intelligence be replaced by a more clearly defined empirical test.

The Turing test measures the performance of an allegedly intelligent machine against that of a human being, arguably the best and only standard for intelligent behavior. The test, which Turing called the imitation game, places the machine and a human counterpart in rooms apart from a second human being, referred to as the interrogator (Figure 1.1). The interrogator is not able to see or speak directly to either of them, does not know which entity is actually the machine, and may communicate with them solely by use of a textual device such as a terminal. The interrogator is asked to distinguish the computer from the human being solely on the basis of their answers to questions asked over this device. If the interrogator cannot distinguish the machine from the human, then, Turing argues, the machine may be assumed to be intelligent.

By isolating the interrogator from both the machine and the other human participant, the test ensures that the interrogator will not be biased by the appearance of the machine or any mechanical property of its voice. The interrogator is free, however, to ask any questions, no matter how devious or indirect, in an effort to uncover the computer’s identity. For example, the interrogator may ask both subjects to perform a rather involved arithmetic calculation, assuming that the computer will be more likely to get it correct than the human; to counter this strategy, the computer will need to know when it should fail to get a correct answer to such problems in order to seem like a human. To discover the human’s identity on the basis of emotional nature, the interrogator may ask both subjects to respond to a poem or work of art; this strategy will require that the computer have knowledge concerning the emotional makeup of human beings.

 

 

14 PART I / ARTIFICIAL INTELLIGENCE: ITS ROOTS AND SCOPE

The important features of Turing’s test are:

1. It attempts to give an objective notion of intelligence, i.e., the behavior of a known intelligent being in response to a particular set of questions. This provides a standard for determining intelligence that avoids the inevitable debates over its “true” nature.

2. It prevents us from being sidetracked by such confusing and currently unanswerable questions as whether or not the computer uses the appropriate internal processes or whether or not the machine is actually conscious of its actions.

3. It eliminates any bias in favor of living organisms by forcing the interrogator to focus solely on the content of the answers to questions.

Because of these advantages, the Turing test provides a basis for many of the schemes actually used to evaluate modern AI programs. A program that has potentially achieved intelligence in some area of expertise may be evaluated by comparing its performance on a given set of problems to that of a human expert. This evaluation technique is just a variation of the Turing test: a group of humans are asked to blindly compare the performance of a computer and a human being on a particular set of problems. As we will see, this methodology has become an essential tool in both the development and verification of modern expert systems.

The Turing test, in spite of its intuitive appeal, is vulnerable to a number of justifiable criticisms. One of the most important of these is aimed at its bias toward purely symbolic problem-solving tasks. It does not test abilities requiring perceptual skill or manual dexterity, even though these are important components of human intelligence. Conversely, it is sometimes suggested that the Turing test needlessly constrains machine intelligence to fit a human mold. Perhaps machine intelligence is simply different from human intelli- gence and trying to evaluate it in human terms is a fundamental mistake. Do we really wish a machine would do mathematics as slowly and inaccurately as a human? Shouldn’t an intelligent machine capitalize on its own assets, such as a large, fast, reliable memory,

THE INTERROGATOR

Figure 1.1 The Turing test.

 

 

CHAPTER 1 / AI: HISTORY AND APPLICATIONS 15

rather than trying to emulate human cognition? In fact, a number of modern AI practitio- ners (e.g., Ford and Hayes 1995) see responding to the full challenge of Turing’s test as a mistake and a major distraction to the more important work at hand: developing general theories to explain the mechanisms of intelligence in humans and machines and applying those theories to the development of tools to solve specific, practical problems. Although we agree with the Ford and Hayes concerns in the large, we still see Turing’s test as an important component in the verification and validation of modern AI software.

Turing also addressed the very feasibility of constructing an intelligent program on a digital computer. By thinking in terms of a specific model of computation (an electronic discrete state computing machine), he made some well-founded conjectures concerning the storage capacity, program complexity, and basic design philosophy required for such a system. Finally, he addressed a number of moral, philosophical, and scientific objections to the possibility of constructing such a program in terms of an actual technology. The reader is referred to Turing’s article for a perceptive and still relevant summary of the debate over the possibility of intelligent machines.

Two of the objections cited by Turing are worth considering further. Lady Lovelace’s Objection, first stated by Ada Lovelace, argues that computers can only do as they are told and consequently cannot perform original (hence, intelligent) actions. This objection has become a reassuring if somewhat dubious part of contemporary technologi- cal folklore. Expert systems (Section 1.2.3 and Chapter 8), especially in the area of diag- nostic reasoning, have reached conclusions unanticipated by their designers. Indeed, a number of researchers feel that human creativity can be expressed in a computer program.

The other related objection, the Argument from Informality of Behavior, asserts the impossibility of creating a set of rules that will tell an individual exactly what to do under every possible set of circumstances. Certainly, the flexibility that enables a biological intelligence to respond to an almost infinite range of situations in a reasonable if not nec- essarily optimal fashion is a hallmark of intelligent behavior. While it is true that the con- trol structure used in most traditional computer programs does not demonstrate great flexibility or originality, it is not true that all programs must be written in this fashion. Indeed, much of the work in AI over the past 25 years has been to develop programming languages and models such as production systems, object-based systems, neural network representations, and others discussed in this book that attempt to overcome this deficiency.

Many modern AI programs consist of a collection of modular components, or rules of behavior, that do not execute in a rigid order but rather are invoked as needed in response to the structure of a particular problem instance. Pattern matchers allow general rules to apply over a range of instances. These systems have an extreme flexibility that enables rel- atively small programs to exhibit a vast range of possible behaviors in response to differ- ing problems and situations.

Whether these systems can ultimately be made to exhibit the flexibility shown by a living organism is still the subject of much debate. Nobel laureate Herbert Simon has argued that much of the originality and variability of behavior shown by living creatures is due to the richness of their environment rather than the complexity of their own internal programs. In The Sciences of the Artificial, Simon (1981) describes an ant progressing circuitously along an uneven and cluttered stretch of ground. Although the ant’s path seems quite complex, Simon argues that the ant’s goal is very simple: to return to its

 

 

16 PART I / ARTIFICIAL INTELLIGENCE: ITS ROOTS AND SCOPE

colony as quickly as possible. The twists and turns in its path are caused by the obstacles it encounters on its way. Simon concludes that

An ant, viewed as a behaving system, is quite simple. The apparent complexity of its behavior over time is largely a reflection of the complexity of the environment in which it finds itself.

This idea, if ultimately proved to apply to organisms of higher intelligence as well as to such simple creatures as insects, constitutes a powerful argument that such systems are relatively simple and, consequently, comprehensible. It is interesting to note that if one applies this idea to humans, it becomes a strong argument for the importance of culture in the forming of intelligence. Rather than growing in the dark like mushrooms, intelligence seems to depend on an interaction with a suitably rich environment. Culture is just as important in creating humans as human beings are in creating culture. Rather than deni- grating our intellects, this idea emphasizes the miraculous richness and coherence of the cultures that have formed out of the lives of separate human beings. In fact, the idea that intelligence emerges from the interactions of individual elements of a society is one of the insights supporting the approach to AI technology presented in the next section.

1.1.5 Biological and Social Models of Intelligence: Agents Theories

So far, we have approached the problem of building intelligent machines from the view- point of mathematics, with the implicit belief of logical reasoning as paradigmatic of intel- ligence itself, as well as with a commitment to “objective” foundations for logical reasoning. This way of looking at knowledge, language, and thought reflects the rational- ist tradition of western philosophy, as it evolved through Plato, Galileo, Descartes, Leib- niz, and many of the other philosophers discussed earlier in this chapter. It also reflects the underlying assumptions of the Turing test, particularly its emphasis on symbolic reasoning as a test of intelligence, and the belief that a straightforward comparison with human behavior was adequate to confirming machine intelligence.

The reliance on logic as a way of representing knowledge and on logical inference as the primary mechanism for intelligent reasoning are so dominant in Western philosophy that their “truth” often seems obvious and unassailable. It is no surprise, then, that approaches based on these assumptions have dominated the science of artificial intelligence from its inception almost through to the present day.

The latter half of the twentieth century has, however, seen numerous challenges to rationalist philosophy. Various forms of philosophical relativism question the objective basis of language, science, society, and thought itself. Ludwig Wittgenstein’s later philosophy (Wittgenstein 1953), has forced us to reconsider the basis on meaning in both natural and formal languages. The work of Godel (Nagel and Newman 1958) and Turing has cast doubt on the very foundations of mathematics itself. Post-modern thought has changed our understanding of meaning and value in the arts and society. Artificial intelli- gence has not been immune to these criticisms; indeed, the difficulties that AI has encoun- tered in achieving its goals are often taken as evidence of the failure of the rationalist viewpoint (Winograd and Flores 1986, Lakoff and Johnson 1999, Dennett 2005).