|
| Home | Forum | Search |
| eNotAlone > Personal Growth |
|
What is AI?
What AI is Artificial Intelligence (AI) is the study of intelligent behaviour (in humans, animals, and machines) and the attempt to find ways in which such behaviour could be engineered in any type of artifact. It is one of the most difficult and arguably the most exciting enterprise ever undertaken by humanity. The difficulty of the enterprise may not be obvious at first, but it has been painfully learned. Sometimes the quest for AI is likened to deep space exploration in its level of difficulty. However, the truth is that it is much harder than that. In space exploration we have, at least, an understanding of the main technical problems. In AI, unfortunately, that is not yet the case. On the other hand, the difficulty is more than compensated for by the potential rewards, both practical and intellectual. In practical terms, AI has already more than justified itself.As we shall see in chapter 2, applications of AI and spin-offs from AI are already shaping our technology and society and will increasingly do so in the future. | ||||||
The intellectual rewards are even more exciting. AI offers (and is now beginning to provide) a scientific understanding of some of the most difficult questions we could ever pose about ourselves and our world.We are at the start of a journey into the most challenging "inner space" — fundamental questions about what it is to be a thinking being at all. Many people find such a scientific quest unnerving. It is unnerving. The possibility that some of our most cherished human attributes might be scientifically explained could be seen as a sort of threat. It is important not to make this threat into some sort of bogeyman that it is not.As we shall see, there are many aspects of human thought that AI has not yet investigated and it may well be that they remain forever uninvestigated. Secondly, the fact that we have, for example, a scientific account of the formation of a rainbow does not detract from its beauty. If we were at some point in the future to provide a scientific account of the working of human creativity, that would not make any of the products of that creativity less beautiful or less interesting. The wonderful products of the human mind are undiminished by explanation. The real root of the threat is probably a predilection for mystery.Humans tend to like a bit of mystery in their stories about the world and particularly like a bit of mystery in how they take their most cherished decisions and come up with their best ideas. At the same time, though, humans are driven to explore. Just as we must constantly strive to find out about the limits of the universe, so we must constantly probe the details of how our intelligence, and that of other animals,works. Intelligence and its many wonderful products are not threatened by any such enquiry.On the contrary, the surprising difficulties involved in getting some apparently simple aspects of intelligent behaviour from machines can only inspire awe at just how wonderful natural intelligence is. Another important consequence of the definition of AI given at the start of this section is the way in which it clearly transcends conventional boundaries of study. It is both science and engineering, since it involves the study of and the building of intelligent behaviour. Indeed work in AI often tends to make a mockery of the supposed division between science and engineering, as it involves finding out through building. Even more radical is the fact that "intelligent behaviour" is to be found in many diverse places. It is in the communication between bees, the movement of prices on the stock exchange, the use of metaphor in Hamlet, and in the working of an automated air traffic control system. If we wish to understand it we must be prepared to pursue it through all these areas and more. This makes complete nonsense of traditional boundaries between arts and science, between engineering and biology, between the individual and society. AI has always been a truly interdisciplinary enterprise, being at the same time both art and science, both engineering and psychology. If this claim seems too extravagant at the abstract level, you should be aware that AI has already produced programs that simulate the rantings of a paranoid schizophrenic, or the birth, breeding and evolution of synthetic creatures. It has produced programs that have discovered new mathematical theorems and programs that can improvise jazz. It has produced programs that can detect fraudulent financial transactions and robots that can clear away empty coke tins from a laboratory. There are programs that paint pictures, programs that perform medical diagnosis, programs that teach, and programs that learn. This is not an attempt to give the impression that AI has already conquered all areas of knowledge. Far from it; these successes are local, cranky, and fail to generalize across other areas. The best that can be said is that they are tantalizing little glimpses of what might one day be achieved.Most importantly, they are evidence of the almost incredible breadth of vision that inspires research in AI.
The first, and perhaps the most important, step in understanding AI is to abandon one's preconceptions.Most readers will have already formed some hazy ideas of what AI is about. These hazy ideas will be almost completely wrong.As far as possible, you should dismiss them before continuing further. You will, for example, have some sort of understanding of the term "intelligence". This might lead you to assume that artificial intelligence has something to do with building the sort of intelligence which you possess into some sort of machine. However, as we shall see repeatedly throughout this book, that assumption can be highly misleading.Work in artificial intelligence keeps on showing that we do not understand our own intelligence in any scientific way. Even more surprising is how it has revealed that the variety of ways in which humans use their intelligence to solve problems are certainly not the only ways available and often they are not the best. There are some good reasons (discussed in chapters 3 and 4) for believing that the study of human intelligence is often not helpful in AI.Not only do we lack scientific understanding of most of the relevant details of human intelligence but it would be way beyond the state of the art to try to imitate it in machines. Many AI researchers look at supposedly simpler animals, such as insects, believing that human intelligence is just too complex to inform their work at all. On the other hand, other AI researchers have achieved spectacular results in getting particular aspects of human behaviour from machines. A good example would be chess playing. Certainly, when work started in this area in the 1950s it was seen as a good example of intelligent human behaviour. In a tournament in 1997 a computer called Deep Blue beat the World Chess Champion, Gary Kasparov. That machines can now play chess very well has been established. However, when we look in detail (in chapter 2) at the way in which computers play chess we shall see that it is probably very different from the ways in which humans play chess. It might seem a little provocative to say that computers play chess better than humans do but, in this field, better means winning. Beating the human World Champion is as much as we can reasonably ask of a chess-playing program. In at least this one area, it seems that we can justifiably talk of having found better methods than those used by humans. Another set of preconceptions about AI comes from the world of science fiction. Intelligent machines, robots, cyborgs and so on are favourite themes of all forms of science fiction. Unfortunately once again, the impressions gained from science fiction are highly misleading. It is important to remember that science fiction is essentially just that — fiction. It has very often inspired scientists in AI and other areas, but it may give a mistaken picture of what is actually happening in current research. In particular, science fiction may lead readers to assume that far more has been achieved than is actually the case and that AI is more human-like than is actually the case. This book will challenge these assumptions. The final set of preconceptions comes from some pervasive myths about computers. These myths are so widely believed, even by some computer scientists, that it requires courage to challenge them, but challenge them we must. It is often said that "computers can do only what we tell them to do". Like all good myths, this one has an element of truth. All computers need elaborate software (usually programs written by humans) in order to operate at all.However if this slogan is read as "all that computers ever do is to follow explicit instructions" then it is very wrong. In this book readers will be introduced to programs that take guesses, play hunches, and in many ways out-perform their creators.We will hear about robots that have not been designed, but rather have been evolved. For similar reasons, readers are implored to relinquish the myth that computers are purely rational, deductive machines. AI researchers have not confined their study and experimentation to the rational parts of intelligent behaviour. It is true that greater success has been achieved in the obviously more rational areas, but AI has also involved getting computers to do some very non-rational things. One discovery of research in AI is that whole areas of intelligent behaviour have nothing to do with logic and deduction. Some different methods are needed, but there have been surprising successes in getting computers to perform in these areas too.
AI researchers have never defined their area in terms of one particular set of research methods.We might say that in AI there are at least as many methods as there are researchers. There are a number of reasons for this. First, such a wide-ranging and crossdisciplinary enterprise simply has to be eclectic in its choice of methods. Second it has always been easy for individual AI researchers to pursue multiple goals. Consider, for example, the field of AI known as NLP (natural language processing). This involves the study of and attempt to build computer programs with which we might communicate in English or other human languages. It has the simultaneous goals of (at least) making computers easier for us to use, understanding the complex rules that govern the structure of natural languages, and of discovering just how humans can learn and apply those rules. Different researchers may give different amounts of emphasis to each of these three goals. Indeed the same researcher may well give different goals prominence on different occasions, depending on the audience. What is true of the area of NLP is true of AI as a whole. It is a study that has always had multiple goals. To extend a crude military metaphor often used in science, we could say that AI has chosen to attack its problem area on the widest possible front. Instead of a concentrated attack that might lead to the core of the problem of intelligence, there are innumerable little skirmishes, along a line that stretches across most of human knowledge. Some of these skirmishes are going well and some not so well, but over a period of a few years that can change. Perhaps understandably, those researchers whose skirmish seems to be yielding results tend to shout that this is the long awaited breakthrough.When the dust settles, however, it becomes clear that once again they have only moved the frontline forward a few yards.At the same time, other researchers may give very plausible reasons why their particular part of the line is where the big push should be made. In the history of AI, however, such "big pushes" have only resulted in moving the line forward locally a few hundred yards. So far, there is steady progress in AI, but no big breakthroughs. The main tool used in AI research is the digital computer. This most certainly does not mean that AI is about digital computers. The computer is, firstly, a tool. It is used because it enables researchers quickly to build "models of behaviour" and examine them. Some AI researchers feel that it is necessary to build real robots that interact with the real world.We shall look in detail at the reasons why they believe this in a later chapter. For now, it should suffice to say that most AI research makes extensive use of computer programs that model aspects of the real world. There are a number of difficult issues raised by this method of research.Many readers will intuitively feel that there is a world of difference between simulated thought and "real thought" and we will go into more detail on this later in the book. For now, it should be sufficient to say that in many other areas this use of computers is very effective.We would expect civil engineers, for example, to make use of computer simulation to discover how a bridge might withstand hurricane force winds. To build a real bridge and wait for "once in a century"weather conditions would be stupid. The digital computer can answer the civil engineers' questions in a matter of hours. It is the ability of the computer to work through myriad possibilities in a very short time that makes it such a useful tool in AI. Modern computers allow us to build detailed working models of things far more complex than bridges and winds. Just as we can now store pictures and music in digital format, so we can work with digital manipulations of anything that we can describe accurately enough. This need for accurate description is, perhaps, the key to understanding AI research methods. The second important role of the computer (or physical robot) in AI is to prompt researchers to ask certain specific questions about natural intelligence. The computer is used not only to simulate, as with the bridge, but also to inspire certain scientific questions. Asking how we might get the computer to do it imposes a new scientific rigour on our way of looking at familiar things. An analogy is often made between AI and "artificial flight". It is an analogy that can be usefully deployed here.At the beginning of the twentieth century there was a very limited understanding of the way in which birds and insects were able to fly: it was obvious that they did and the scientific enquiry tended to end at that point. Indeed, at the time when the Wright brothers started the aviation age (in 1903) most biology textbooks said that birds could fly "because they had the power of flight". This is an account that owes much to Aristotle, writing in Athens in the fourth century BC.However, it is not much help to those who wish to understand how to build aircraft.As a result of the more detailed understanding of flight which stemmed from the building of successful aircraft, we now know that birds fly by complying with the laws of aerodynamics. The quest for some sort of analogous "aerodynamics of intelligence" is often seen as the ultimate goal of AI. Just as scientific understanding of bird flight stemmed from the building of aircraft so, it is hoped, scientific understanding of intelligence will stem from attempting to build intelligent machines. Of course, AI research involves far more than just writing computer programs. If you wish to build something resembling the intelligence of a particular animal then it behoves you to study that animal in great detail.Much AI research involves doing a sort of biology or a sort of psychology, and even a sort of philosophy. Ignoring fatuous "turf wars" between disciplines (and such wars deserve nothing more), we can see that the use of the computer as a tool makes these "sorts"more rigorous than they would be without it. An illustrative example is not hard to find. Take the task which you are now performing: reading.Obviously you have "the power of reading" to have got this far. That you have reasonably clear printing to follow and at least one suitable eye (or fingertip, if you read in Braille), we might also deduce.However, these truths would not be of much help in the design of a reading machine. If we wish to build a machine that can read we need to ask some more detailed questions. Are you, for example, looking at each individual letter and comparing it to a library of characters (at least 114 in number) to determine what it is, before proceeding to the next? On reaching a blank space and assembling the letters into a word, do you then search through some sort of dictionary (a much larger number is involved here) in order to recognize it? After all this searching and decoding you will still need to assemble the words into a sentence and extract some sort of meaning. To perform all these tasks in a reasonable period of time (i.e. today, rather than next week) is way beyond even the most powerful computers. On the other hand, you could be using your knowledge to generate "expectations". Knowledge of the rules of English grammar will tell you, for example, that most correct English sentences are composed of a noun phrase and a verb phrase. Finding the verb phrase is (there it is!) the key to understanding the sentence. Even if you have not studied the formal rules of English grammar you probably are, and have been, using them to read.Unfortunately, grammar alone does not enable you to understand a sentence. Knowledge of the world seems to be an important element in being able to extract the meaning from a sentence. It can also tell you in advance what sorts of words and sentences might be coming next.Using world knowledge to reduce the amount of computation involved in reading solves one immediate problem, but at a high price.Now we need to find out how you get that world knowledge and how you can access it so quickly in order to be able to read this. Getting a machine to do this is not going to be easy! Please don't become so self-conscious about your reading that you don't go any further.What is important here is that you don't need to ask these questions to be able to read, but you do need answers to most of them in order to build a reading machine or to give a scientific account of just how it is that you can read. AI makes us at least consider how we might begin the task of building a machine to do something, even if no such machine is actually in prospect. That, in turn, imposes far more intellectual and scientific rigour on the way we look at examples of intelligent behaviour. Even if the machine we consider is just a theoretical machine that we can't yet build, the discussion must get beyond superficial accounts of how we do things. Just as the crudest of aircraft ended the "power of flight" approach to bird flight, so consideration of even the crudest intelligent machines means that we must look at biology, psychology, and linguistics in a much more rigorous and detailed way.
As we have seen, AI involves a tremendously wide range of problems and approaches. In fact this range is so wide that it often seems to some AI researchers that others in the field are not even doing the same subject. In practice this is not usually a problem."Let a thousand flowers bloom" is a popular slogan in AI. For some commentators, on the other hand, it is important to be able to give some account of the "ultimate goal" of the subject. One research team, for example, might be spending their time machining precision gear-wheels with the goal of producing a robot that can walk up stairs without falling over. Another team might be analyzing literature to see if they can determine any rules underlying the use of metaphor. Their ultimate goal might be a computer program that can recognize and respond to metaphors in human input.How can we say that they are, in some real sense, engaged in the same project? Saying just what is the final goal that unites such different strands of research is not easy and wrong answers may have unfortunate consequences. Over the history of AI, there have been a number of attempts to make a single simple description of the ultimate goal of AI and they have generally been unsatisfactory. Many contemporary researchers would prefer not to confront the problem. They prefer instead to concentrate on their own local goals.However, in doing so they may be missing the sort of interdisciplinary crossover that has often proved so useful in AI. The gear-wheel machinists may need to know, for example, how insect legs are articulated, what birds need in order to be able to balance on two legs and so on. The literature analyzers may need to know about work in multi-valued logic which might bear on the ways in which metaphors work. In spite of all the difficulties, it is still worth looking at some of the suggested answers to the question "what is the ultimate goal of AI?" We have already seen one possible approach and my personal favourite is the "aerodynamics of intelligence"mentioned in the preceding section. This approach would claim that the ultimate goal of AI is to produce a full scientific account of human, animal, and machine intelligence showing the common principles underlying all three. The problem with this, it must be admitted, is that we know very few, if any, of those common principles at the moment.We will look at this in more detail in chapter 5. Many of the other approaches to defining the ultimate goal of AI have tended to stress the development of "humanlike" levels of intelligence in machines. These need to be treated with caution for the reasons we have already seen.However, one of them — the so-called "Turing test" has had such a tremendous influence on the history of AI that we must spend the next two sections examining it in detail.
Undoubtedly the most famous answer to the question "what is the ultimate goal of AI?" is provided by the so-called Turing test. I say "so-called" because Alan Turing, after whom it is named, never talked of a test. Indeed, there are so many misinterpretations of Turing and the test attributed to him that it might accurately be called AI folklore.With your indulgence, I would like therefore to relate the whole story in some detail. Alan Turing was undoubtedly a genius. Shortly after graduating in mathematics at Kings College, Cambridge he wrote a paper (published in1936) which revolutionized our understanding of the nature of mathematics. That would have been enough for the average genius, but for Turing it was only the beginning.During World War II (in September 1939 to be precise) he and a number of high-powered intellectuals were secreted away by the British military in a requisitioned stately home called Bletchley Park. This is now part of a suburb of Milton Keynes in the South of England and is well worth a visit. There they worked on breaking German military codes known as Enigma. In this objective they were supremely successful. Turing himself played a central role in understanding how Enigma was in fact breakable. That the code could be broken was never considered even possible by the Germans. Even at the end of the war, when it was obvious that the Allies had advance knowledge of German movements, the German high command was searching for traitors rather than considering the possibility that the Enigma code might have been cracked. It should be obvious that the British ability to decode most German secret transmissions was a war-winning advantage. Even the most conservative historians admit that this achievement shortened the war by at least a year. Less obvious is the fact that it all remained highly secret. In fact, the truth about what had happened at Bletchley Park did not begin to emerge until the 1980s, and some aspects remain classified even now.Most importantly for present purposes, the code-breaking work at Bletchley Park involved the use of machines which were the precursors of modern computers. The Enigma code was so called because it was generated and decoded by the Enigma machine. Various other machines were used by the British code breakers. The most important of these was known as Colossus. This had most of the features of modern electronic computers, but in the rather foolish desire for total secrecy all ten of these machines at Bletchley Park were totally destroyed at the end of the war. This left Turing and his colleagues in an embarrassing position. They knew then enough to build effective electronic computers, but they could not really say how they knew. The fact that they had seen such machines working day in and day out at Bletchley Park could never even be hinted at. Eventually a small team built a machine at Manchester University. It is from this machine that all modern computers are descended. In 1948 Alan Turing was writing programs for this machine. He was also writing a paper entitled "Computing Machinery and Intelligence". This paper laid out the ideas which became known as the Turing test. "Computing Machinery and Intelligence" was published in 1950 in Mind — one of the longest established British philosophy journals. Let us note that it was a paper written by a mathematician, turned code-breaker, turned computer programmer, and was published in a philosophy journal. The interdisciplinary nature of AI has been apparent since its very outset. In "Computing Machinery and Intelligence"Turing says he wishes to discuss the question, "Can machines think?"However, since this question is too vague, he proposes replacing it with a game. This game he called the "imitation game". It involves three people in separate rooms. They can communicate only by typing messages to each other. In the original version, there is a man, a woman, and an interrogator whose gender is unimportant. The interrogator, as the name suggests, can ask any question of the other two participants. The objective of the game is for both the man and the woman to convince the interrogator that they are the woman. The woman will be answering truthfully and the man will be typing things like "Don't listen to him, I'm the woman". (It's rather like what goes on in some Internet chatrooms.) Now what would we say, asks Turing, if the role of the man in this game were to be successfully played by a machine? That is if, after five minutes of questions, the average interrogator would not be able to recognize that he or she was communicating with a machine on at least thirty percent of occasions. If we could make machines that could do this well in the imitation game, then ordinary people would be happy to say that they were thinking machines. In fact, Turing thought that it was a matter of "when", not "if ", we would make such machines.He confidently predicted that, by the year 2000, digital computers would be able to achieve this level of success in the imitation game. This achievement would change public attitudes so that it would become normal to talk of "thinking machines". One remarkable point about this paper is that Turing managed accurately to predict the level of computing power that would be available by the year 2000. This was despite the only real contemporary example being the Manchester machine whose roomful of equipment had much less computing power than we get from a tiny microchip nowadays. Turing was right in his prediction about the growth of computer power; however, no computer is anywhere near good enough to succeed in the imitation game in the foreseeable future. Unfortunately, Alan Turing's career ended not long after the publication of this paper.He committed suicide in 1954, still only 42. One interesting final point about his life is that during the 1950s his interests and publications had moved on to the mathematical foundations of biology — an area that began to excite AI researchers again in the 1990s.
There are a number of reasons why the Turing test should not be seen as any sort of goal for AI, least of all the ultimate goal. Firstly, it concentrates on human performance and that is an unnecessary restriction for AI. AI is also concerned with other animals, most of which could not participate in the "imitation game". Secondly, in the actual building of machines, it is a distraction constantly to have to imitate human methods and performance. Some people in AI do not agree with my last paragraph. Indeed, candidate programs are still entered into an imitation game competition every year. This is known as the "Loebner prize" after Dr Hugh Loebner, the inventor and industrialist who has offered a prize of 100,000 dollars for the first computer program to pass his version of the Turing test. Although no program has yet won the Grand Prize, there is a smaller prize of 2000 dollars for the most human-like computer program in the competition each year and this attracts a number of good attempts. Looking at these programs in detail reveals another problem with treating the Turing test as the ultimate goal of AI. All of the 2000 dollar prize-winners have been fairly simple computer programs that are designed to give the illusion of holding a conversation. Nowadays such programs are called "chatbots". They have a number of set responses which they print out in response to various inputs by the interrogator. This is a technique which was first used in a program called ELIZA in 1966. The name is taken from Eliza Dolittle in Shaw's play Pygmalion. It is not used entirely accurately, however, since Eliza Dolittle was taught to speak, while the eponymous program merely gives the illusion of being able to speak. There has been some refinement of such illusion programs since 1966, but these refinements do not really contribute to progress in AI. For example, if the program prints out text that is highly opinionated about matters that are political or sexual, then interrogators are more likely to think that it is human. This tells us something about human psychology which might be of some interest, but it tells us nothing about how to build intelligent machines. This reveals a third serious problem with treating the Turing test as the ultimate goal of AI. It pushes researchers to produce programs that are primarily aimed at deceiving humans, not at any more fundamental approach to the problem of intelligence. In the next chapter we will look at some programs for which the claim of passing the Turing test has been made. It should quickly become obvious that they are actually more about deceit than about intelligence. Many people working in AI would agree with my criticism of the Turing test, but say that it is still relevant because a truly intelligent machine (whatever that means) would be able to pass the test — mainly as a by-product of being intelligent. This may or may not turn out to be the case, but we are unlikely to see such a machine built in the near future. Remember that the interrogator may ask absolutely any question. This makes the Turing test a very hard test indeed. Realistically, it is extremely difficult, expensive, and ultimately pointless to set about the project of building a machine that could pass the Turing test. Direct mimicry of human intelligent performance is unlikely to prove profitable when there is plenty of human intelligence available.
|
| |||||
|
© 2008 eNotAlone.com | ||||||