APPERANCES OF ARTIFICAL INTELLIGENCE

                           Appearances of AI

The degree to which machines appear intelligent first, hinges on whether or not the work they perform is intellectual (e.g., doing arithmetic) or manual (e.g., slicing steaks): here, an electronic calculator is a more suitable candidate than an electric carving knife. A second is the degree to which the machine is self-actuated (self-propelled, self-activated, and self-controlled), or "autonomous": here, an electronic calculator is a stronger contender than an abacus. Computers are stronger contenders than calculators on both lines. Where old-style AI aims to raise computer intelligence quotients (so to speak), nouvelle AI aims at facilitating robot autonomy.

a. Computers

i. Prehistory

Initially, tools (such as axes) were human physical powers extended; initially powered by human muscle; then by domesticated animals and in-place forces of nature, including water and wind. The steam engine placed fire in their bellies; machines were self-powered, vestiges of self-governance (as by Watt's 1788 centrifugal governor); and the rest is modern history. Automation of intellectual labor meanwhile had commenced. Blaise Pascal created an early adding/subtracting machine, the Pascaline (around 1642). Gottfried Leibniz incorporated multiplication and division capabilities with his Stepped Reckoner (around 1671). The first programmable device, however, worked on fabric not numbers. The Jacquard loom invented (c. 1801) by Joseph-Marie Jacquard employed a punched-card system to mechanize the weaving of patterns and designs programmable by machine: in a dramatic demonstration, the loom was set to weave a silk tapestry portrait of Jacquard himself.

In plans for his Analytical Engine mathematician/inventor Charles Babbage noted (around 1836) that the punched cards could direct operations on symbols as easily as they could on silk; the cards could represent numerals and other symbolic information and, more significantly, instructions, including conditionally branching instructions, for numeric and other symbolic operations. Augusta Ada Lovelace (Babbage's computer programmer) understood the significance of these advances: "The limits of arithmetic" she says, "were … transcended the instant the notion of using the [instruction] cards had struck" thus "enabling mechanism to be associated together with general symbols, in successions of unlimited variety and extent" (Lovelace 1842). "Babbage," Turing observes, "had all the necessary ideas" (Turing 1950). Babbage's Engine – if he had built it in all its steam powered cog-wheel driven splendor - would have been an all-purpose programmable device, the first computer. 

ii. Interlude: Turing Machines

Prior to the possibility of automated computation with the invention of electronic computers in the mid twentieth century, Alan Turing established the theoretical basis of Computer Science by defining with clarity the connection Lady Lovelace predicted "between the operations of matter and the abstract mental processes of the most abstract branch of mathematical sciences" (Lovelace 1842). Turing (1936-7) explains a kind of machine (now known as a "Turing machine") which would be able to compute any conceivable algorithm, or carry out any "rote" task. Since Alonzo Church (1936) – employing recursive functions and Lambda-definable functions – had established the very same collection of functions as "rote" or algorithmic as Turing machines could calculate, this significant and widely accepted identification is referred to as the "Church-Turing Thesis" (see, Turing 1936-7: Appendix). The machines Turing described are

only able to be in a finite number of states … "m-configurations." The machine is provided with a "tape" (the equivalent of paper) passing through it, and segmented into sections (referred to as "squares") each able to carry a "symbol." At any given time there is only one square … which is "in the machine." … The "scanned symbol" is the only one of which the machine is, so to speak, "directly aware." But by changing its m-configuration the machine can effectively remember some of the symbols which it has "seen" (scanned) before. The machine's possible behavior at a given moment is specified by the m-configuration … and by the symbol being scanned …. This pair … the "configuration" … specifies the possible behaviour of the machine. In certain of the configurations of which the square is empty … the machine writes a new symbol on the scanned square: in others it deletes the scanned symbol. The machine can also transform the square that is being scanned, but only by moving it one position to right or left. Besides any of these operations the m-configuration can be altered. (Turing 1936-7)

Turing then demonstrates how such machines can represent actionable descriptions of other such machines. Thus, "It is possible to invent a single machine which can be used to compute any computable sequence" (Turing 1936-7). Present-day digital computers are (and Babbage's Engine would have been) embodiments in physical material of this "universal computing machine" that Turing formulated abstractly. Theoretically, this implies that anything that can be done algorithmically or "by rote" at all "can all be done with one computer suitably programmed for each case"; "considerations of speed apart, it is unnecessary to design various new machines to do various computing processes" (Turing 1950). In theory, whatever their architecture or hardware (below), "all digital computers are in a sense equivalent": equivalent to Turing's "universal computing machine" in speed-apart capacity.

iii. Practice

Practically speaking, where speed is not apart, hardware and architecture matter: the more speedy operations the higher the computational capacity. As much as innovation on the hardware front from cogwheels to circuitry was necessary to make digital computers feasible at all, advancements in computer performance have been dependent to a large extent on ongoing innovation in making faster, more and more capable, machines. Electromechanical relays were replaced by vacuum tubes, tubes by transistors, and transistors by increasingly integrated circuits, providing enormously enhanced operation speeds. In the meantime, memory has become faster and less expensive.

Architecturally, with the exception of the earliest and some subsequent experimental machines, they have a stored program serial design commonly referred to as "von Neumann architecture" (named after John von Neumann's contribution to the design of EDVAC, the first computer to store programs as well as data in working memory). The architecture is serial in the sense that operations are executed sequentially by a central processing unit (CPU) that has a rich repertoire of elementary operations: even so-called "reduced instruction set" (RISC) chips have basic operation sets much richer than the bare minimum few Turing demonstrated theoretically necessary. Parallel structures, on the other hand, spread out computational operations among two or more units (usually many more) that are able to function simultaneously, each with (maybe) severely limited basic operating capabilities.

Gordon Moore, one of Intel's founders, noticed in 1965 that transistor density on integrated circuits had doubled annually since 1959 when they were invented: "Moore's law" is used to predict continuing similar rates of exponential growth in chip density (specifically), and computer processing power (in extension), well into the future. Advances on the programming of software – while necessary and by no means irrelevant – have appeared stuttering by comparison. The journey from power to performance is turning out more bumpy than Turing had hoped. Still, machines today do act in numerous ways that would be described as intelligent in animals and humans. Currently, machines perform many activities previously only done by animals and believed to constitute some form of intelligence in them, such as seeking, discovering, and monitoring things; appearing evidence of low-level AI. Currently, machines also perform things previously only accomplished by humans and believed to represent high-level intelligence in humans; for instance, performing mathematical discoveries, playing board games, planning, and learning; appearing evidence of human-level AI.

b. "Existence Proofs" of AI

i. Low-Level Appearances and Attributions

The activities of most machines – some a good deal less sophisticated than computers – lead us to characterize them in mental terms generally reserved for animals. Some missiles, for example, hunt heat, or so we put it. We refer to them as "heat seeking missiles" and no one minds. Room thermostats sense room temperatures and attempt to maintain them at specified ranges by cycling the furnace on and off; and if you place dry ice near its sensor, it will read the room temperature as lower than it is, and erroneously heat the room (see McCarthy 1979). Seeking, monitoring, attempting, and assuming things to be the case appear to be mental states or processes, characterized by their intentionality. Just as humans share low-level mental properties – like seeking and perceiving things – with the lower animals, so too do computers appear to share such low-level properties with less complex devices. Our operational descriptions of computers are full of low-level mental ascriptions: we attribute to them detecting key presses, attempting to initialize their printers, looking for available devices, and so on. Even people who would refuse the statement "machines think" when set down in so many words, are irresistibly driven in their everyday transactions to describe the activities of computers in cognitive terms, and would find it difficult otherwise. In this regard, Turing's prophecy that "at the end of the century the use of words and general educated opinion will have altered so much that one will be able to speak of machines thinking without expecting to be contradicted" (Turing 1950) has been as mightyly realized as his prophecy of some machine success at playing the Imitation Game has been refuted. The Turing test and AI as classically conceived, however, are more concerned with high-level appearances such as the following.

ii. Theorem Proving and Mathematical Discovery

Theorem proving and mathematical exploration being their home turf, computers have displayed not only human-level but, in certain respects, superhuman abilities here. For speed and accuracy of mathematical calculation, no human can match the speed and accuracy of a computer. With regard to high-level mathematical achievements, i.e., theorem proving and mathematical discovery, a start was made by A. Newell, J.C. Shaw, and H. Simon's (1957) "Logic Theorist" program that successfully proved 38 of the first 51 theorems of B. Russell and A.N. Whitehead's Principia Mathematica. Newell and Simon's "General Problem Solver" (GPS) carried similar computerized theorem proving technology beyond the very limited domain of pure logic and mathematics. Today these techniques are in common use in expert systems such as MYCIN, in tutorial logic software, and in programming languages such as PROLOG. There are even new mathematical findings due to computers. In particular, K. Appel, W. Haken, and J. Koch (1977a, 1977b), and computer, demonstrated that all planar maps are four colorable – a significant mathematical conjecture which had defied unaided human proof for more than a hundred years. Some of the computer generated aspects of this proof are so intricate that they cannot be verified (without a computer) by human mathematicians.

While attempts to use general reasoning on unlimited domains are inhibited by inferential explosiveness and the lack of common sense in computers, expert systems circumvent these issues by limiting their domains of application (effectively, to microworlds), and designing domain-specific inference rules for these bounded domains. MYCIN, for example, uses rules derived from interviews with experienced human diagnosticians to descriptions of patients' presenting symptoms to diagnose blood-borne bacterial infections. MYCIN exhibits diagnostic ability near the expert human level but strictly confined to this narrow area. Fuzzy logic is a formalism for encoding vague concepts like most and baldand and allowing inferences based on such facts as that a bald individual predominantly lacks hair.

iii. Game Playing

Playing games absorbed the attention of AI researchers from nearly the beginning. Samuel's (1959) checkers (or "draughts") program was impressive in that it incorporated mechanisms for learning from experience sufficiently so that it could, after a while, outplay Samuel himself. Also, in having one version of the program play against a slightly different version, passing the settings of the better player to the next generation, and doing it again and again – allowing stronger and stronger versions to develop – Samuel developed the application of what have been termed "genetic algorithms" and "evolutionary" computing. Chess has also prompted important achievements leading, in 1997, to Deep Blue's landmark defeat of world champion-in-defence Gary Kasparov in a well-covered sequence of games (chronicled in Hsu 2002). Although some in AI criticized Deep Blue's use of "brute force" computing rather than more refined search directing heuristics, we can still include chess in checkers (where the incumbent "human-machine machine champion" since 1994 has been CHINOOK, the computer) and backgammon as games against which computers now play at or near their highest human levels. Computers play fair to middling poker, bridge, and Go – albeit not at human top level. Moreover, intelligent agents or "softbots" are components or players in all sorts of computer games.

iv. Planning

Planning, largely, is what places the mind in intellectual games such as chess and checkers. To enable this wider mental capacity to be automated was the goal of Newell and Simon's General Problem Solver (GPS) program. GPS managed to solve problems such as the cannibals missionaries problem (how to ferry three missionaries and three cannibals across a river in a canoe for two so that the missionaries are never outnumbered on either bank) by "establishing subgoals whose achievement leads to the achievement of the [final] goal" (Newell & Simon 1963: 284). By these techniques GPS would "generate a tree of subgoals" (Newell & Simon 1963: 286) and attempt a path from initial state (e.g., all on the near bank) to final goal (all on the far bank) by heuristically guided search along an expanding "tree" of potential actions (e.g., two cannibals cross, two missionaries cross, one of each cross, one of either cross, in either direction) until it discovers such a path (e.g., two cannibals cross, one returns, two cannibals cross, one returns, two missionaries cross, … ), or else discovers that there is none. Because the number of branches grows exponentially with the number of options presented at every step, where paths consist of many steps with many alternatives presented at every choice point, as in the real world, combinatorial explosion occurs and a complete "brute force" search is computationally impractical; therefore, heuristics (imperfect rules of thumb) for recognizing and "pruning" the most unfruitful branches so as to allocate more focus to promising ones are required. The widely used STRIPS formalism originally developed at Stanford for Shakey the robot in the late sixties (see Nilsson 1984) describes actions as state operations, each operation with preconditions (described by state descriptions) and effects (described by state descriptions): for instance, the go(there) operation may have the preconditions at(here) & path(here,there) and the effect at(there). AI planning methods are increasingly being applied and even becoming essential in a wide variety of intricate planning and scheduling problems such as airport arrivals, departures, and gate allocation; store inventory control; automated satellite missions; military supply chain; and numerous others.

v. Robots

Sense-model-plan-act (SMPA) approach-based robots led by Shakey, however, have been slow to emerge. Even though it worked in a simplified, tailor-made experimental setting or microworld and used the most powerful offboard computers available, Shakey "ran agonizingly slowly" (Brooks 1991b), as have other SMPA based robots. A paradoxical discovery of robotics research is that skills like object recognition and obstacle avoidance shared by humans with "lower" animals tend to be more challenging to realize than uniquely human "high level" mathematical and inferential skills that are more naturally (so to speak) suited to computers. Rodney Brooks' alternative behavior-based methodology has been successful in teaching low-level behavioral skills outside of specially designed microworlds, but it is difficult to envision how such an approach could ever "scale up" to facilitate high-level intelligent action (see Behaviorism: Objections & Discussion: Methodological Complaints). Maybe hybrid systems can do better than the limitations of either approach. On the technical side, however, things are moving ahead: NASA's Mars exploration rovers Spirit and Opportunity, for example, included autonomous navigation capabilities. If space is the "final frontier" the final frontiersmen are likely to be robots. Meanwhile, Earth robots appear set to get smarter and more ubiquitous.

vi. Knowledge Representation (KR)

Knowledge representation encapsulates concepts and information in computationally available and inferentially manageable forms. In addition to the STRIPS formalism discussed above, some of the other significant knowledge representation formalisms are AI programming languages like PROLOG, and LISP; data structures like frames, scripts, and ontologies; and neural networks (below). The "frame problem" is the problem of faithfully updating dynamic systems' parameters when other parameters change so as to express commonsense generalizations: that things' colors are not changed by their being relocated, that things' locations are not changed by their being colored, and the like. More effective representation of commonsense knowledge is generally believed to constitute a key obstacle to the creation of the kind of interrelated planning and thought processes characteristic of human or "general" high-level intelligence. The CYC project (Lenat et al. 1986) at Cycorp and MIT's Open Mind project are persistent efforts to construct "ontologies" of commonsense knowledge in computer useful forms.

vii. Machine Learning (ML)

Learning – improvement in performance, concept formation, or acquisition of information through experience – underpins human common sense, and one wonders if any preformed ontology ever could give common sense in full human measure. In addition, whatever the other intellectual talents something may exhibit (or appear to), at whatever elevated a degree, without learning ability, it would still appear to lack something essential to human-level intelligence and maybe intelligence of any degree. Machine learning is implicit in computer programs' capacity for self-alteration and several ways of making that possible keep being crafted. The forms of machine learning methods include ensemble learning, explanation-based learning, current-best-hypothesis learning, instance-based learning, reinforcement learning, neural networks, Inductive Logic Programming (ILP), decision tree learning, Bayesian statistical learning, and Bayes networks. All these have seen a variety of uses ranging from game programs that learn through play to data mining (extracting patterns and patterns of information).

viii. Neural Networks and Connectionism

Neural or connectionist networks – made up of simple processors or nodes operating in parallel – are intended to more closely simulate the structure of the brain than conventional serial symbol-processing systems. Assumed brain-computations would appear to be carried out in parallel by the operations of countless brain cells or neurons. Just as their parallel processing is distributed across multiple, possibly dispersed, nodes, data representation within such connectionist networks is likewise distributed and sub-symbolic (not being phrased in formalisms like conventional systems' machine codes and ASCII). Proficient at recognizing patterns, these networks appear especially well-suited to constructing concepts autonomously on the basis of feedback from experience and possess several other humanoid cognitive features in addition. Whether neural networks are capable of using high-level symbol processing like that required in generating and understanding natural language has been vigorously debated. Critics (e.g., Fodor and Pylyshyn 1988) claim that neural networks are, in principle, unable to realize syntactic structures that are sufficient for compositional semantics – where the meaning of larger expressions (e.g., sentences) is composed from the meanings of constituents (e.g., words) – such as those natural language comprehension properties. Conversely, Fodor (1975) has contended that symbol-processing systems cannot acquire concepts: in this case, the pattern recognition abilities of networks appear to be just the job. In this case, as with robots, hybrid systems can perhaps beat both the parallel distributed and symbol-processing methods at their own game. 

ix. Natural Language Processing (NLP)

Natural language processing has turned out more recalcitrant than one would have predicted. Languages are symbol systems and (serial architecture) computers are symbol crunching machines, each with their own proprietary instruction set (machine code) into which they translate or compile instructions phrased in high level programming languages such as LISP and C. One of the main challenges presented by natural languages is the correct assignment of meaning. High-level computer languages articulate imperatives that the machine "understands" procedurally through translation into its native (and also imperative) machine code: their structures are essentially instructions. Natural languages, by contrast, possess – perhaps most importantly – declarative purposes: their structures contain descriptions whose meaning appears essentially to involve rightly relating them to their referents in the world. In addition, computer language at the high level has distinct compilations of machine codes (in a specific machine), while, the identical constructions in natural languages can have differing senses in distinct linguistic and extralinguistic situations. Compare "the child is in the pen" with "the ink is in the pen" wherein the first "pen" will have to be interpreted as one type of enclosure and the second "pen" one type of writing instrument. Commonsense, briefly, is the way we are certain of this; but how would a machine be certain, unless we were able to equip machines with commonsense? Sophisticated and integrated syntactic, morphological, semantic, pragmatic, and discourse processing, in more than a word, would be required. Though the holy grail of complete natural language understanding is still a distant prospect, here as elsewhere in AI, incremental progress is being achieved and put to use in grammar checkers; information retrieval and information extraction systems; natural language interfaces to games, search engines, and question-answering systems; and even partial machine translation (MT).

c. On the Behavioral Evidence

Low level intelligent action is ubiquitous, ranging from thermostats (to mention a low tech. example) to voice recognition (e.g., in automobiles, cell-phones, and other appliances that respond to verbal spoken commands) to fuzzy controllers and "neuro fuzzy" rice cookers. Everywhere nowadays there are "smart" devices. High level intelligent action, as currently found in computers, however, is episodic, isolated, and disintegral. Artifacts whose intelligent activities would exemplify human-level comprehensiveness, attachment, and integration – like Lt. Commander Data (of Star Trek the Next Generation) and HAL (of 2001 a Space Odyssey) – are still science fiction, and will very likely continue to be so for the foreseeable future. Specifically, the challenge of the Turing test remains unsolved. Whether it ever will be met remains an open question.