The New Thinking Machines

Does thought have to be an exclusively human activity? Ironically,every stage of the debate has helped artificial intelligence become a reality.

Build an artificial intelligence, a computer that thinks like a person, and - as every reader of science fiction knows - it soon wants to take over the world. The reality, of course, is very different. Just as robot-builders have dismally failed to reproduce a whole human body, so researchers in artificial intelligence (AI) have created only fragments of human intelligence. These creations are far stranger than the destructive artificial minds of science fiction, but also far more useful.

'By the end of the century one will be able to speak of machines thinking without being contradicted' Alan Turing, 1950

The irony of artificial intelligence is that tasks that are hard for people - like playing good chess, solving logic problems and juggling complex sets of rules - are relatively easy for machines. But common-sense tasks - like getting to work, holding a conversation, or avoiding danger - are still far beyond the capability of even the smartest machines.

Artificially intelligent computers can beat a grandmaster at chess, but would sit pondering their next move as the room burned down around them. In the factory, artificial intelligences manage production schedules incomprehensible to people - yet they couldn't get to work in the moming if they had to. Computers "read" and route paperwork in banks and offices, but cannot understand a child's first book.

Are such machines "intelligent" then? Certainly they fail by the criterion proposed by computer pioneer Alan Turing. Intelligence,he reckoned,is in the eye of the beholder. A machine could be deemed intelligent only if it could hold a conversation with a person (through a telex machine if necessary, to disguise the fact that the machine lacked a body) and convince the person that it was human. No machine has passed that test and none looks likely to for quite some time.

Make your move, robot
Chess is hard for humans to play well, but machines can beat most people easily. This robot is capable of dismissing a good club player using only a few seconds of computing time. Yet, despite its humanoid shape, it is less able to move about than a young toddler. So how intelligent is that?

But it is precisely because machines do not replicate human abilities that they are so potentially useful. After all, there are plenty of people and, should a shortage arise, there is a well-known technique for creating more.
In today's information economy, any unusual reasoning ability is valuable. Although the extraordinary skills of computers in some sorts of reasoning are offset by equally extraordinary stupidity in other areas, by building the right partnership between human and machine it is possible to offset computer stupidity with human intelligence, and vice versa.

At the beginning of each era of its historical development, Al has thrown up at least one good new idea about the underlying roots of intelligence. By the end of the era the idea is found to be at best only a partial solution - but a partial solution with some potentially useful applications nevertheless.

The first stirrings of artificial intelligence

The first era started in the late 1950s, when Herbert Simon, a professor at Pittsburgh's Carnegie Mellon University (who was later to win a Nobel prize for economics), optimistically named his "General Problem Solver".
Simon and his colleague Alan Newell had spent hours getting students to talk aloud as they solved simple reasoning problems. By analysing the mutterings, they reckoned they had found the general principles underlying intelligence: in short, they thought it was all about finding ways to sift through all the known possibilities to arrive at the right answer.

Smart machines go to the movies

Expert systems hold the nuclear balance
In the early 1970s film The Forbin Project,both America and the USSR entrust their strategic defence to intelligent computers. Naturally, the two machines gang up on the human race.
Artificial Intelligence on a bad day The sentient computer in 2001: A Space Oddysey, HAL (centre), is driven mad by conflicting orders. Today, it is more likely for a human not to understand an AI than the other way around.
Stand up for robot rights Robbie (right) was star of the 1950s sci-fi classic Forbidden Planet. Able to drive a car, serve drinks and hold a pleasant conversation, he would easily have passed the Turing test. Although he opened shops and had his own series, Robbie was still a man in a suit. We are never likely to see his like.

Simon and Newell believed a machine could carry out this procedure, which they called "means-end analysis". If a mechanism could generate, step by step, every possible solution, it would effectively reproduce intelligence.
For their machine to solve a problem, it first had to figure out how to know when it had found the answer. One type of problem studied by Simon and Newell was the cryptarithmetic puzzle, for example "SEND + MORE = MONEY". You've got the solution when you know which digits substitute for which letters in the equation.

Machines adapt to changes of plan easily - but they fail other tests that humans pass

Planning by computer means you can adapt to any change instantly. If a shipment is held up at the docks (above) the computer can reschedule production. The same AI system can make sure best use is made of warehouse space (top)

Then the General Problem Solver would look for means to achieve that end. It might begin by trying to substitute one for S and see if that could be the first step towards a solution. If it wasn't, the machine would try two, and so on. Having tried out all the means, the machine selects only the ones that meet its goals.

The problem is that it can, and typically does, take so long for the computer to find a solution as to render it useless. A computer searching for a way to get to work on time is wrong if arrives at lunchtime - even if it has found a more cunning way than any previously devised. And, as can be shown mathematically, there are some problems that would require at least the age of the known universe for a computer to solve by the means-end method.
There are, however, techniques for narrowing down the search. In our cryptarithmetic problem, for example. even a fairly simple- minded computer could be programmed to spot that the letter to start with is M, as it is the first digit of two of the terms, one of which is the answer. (It must equal one as, no matter what each digit is, two of them added together will not equal more than 18: so only a one will ever be carried over.)

For a few problems, a clever search will outdo even the best human performance. Chess is probably the best example. The computers that now routinely beat grandmasters have no grasp of strategy at all. Unlike humans, they can analyse several million possible chess positions a second, and score each according to some simple measure. In this way the computer can look six, eight, even ten moves into the future. In chess, the key to narrowing the search is to assume that the opponent will always make the best move available. Once the computer finds the replies that would leave it worse off, it can dismiss any move that allows the opponent to make those replies. This facility for sifting through vast realms of possibilities for the best combination of steps towards a given goal gives machines many planning applications. Computers helped with the logistics of moving men and equipment into place for the Gulf war. Another military use for AI is organising maintenance schedules for aircraft. Black and Decker uses Alto schedule its production of tools.

The big advantage of planning by computer is the ability to keep track of interactions between the various parts of a production plan. If, say, a shipment of parts is delayed, the computer can quickly reschedule production to make sure machines do not sit idle. Despite their successes at chess and planning, today's computers cannot solve most other kinds of problems. Computers cannot play the Japanese game Go, for example, because at each turn there are many more moves available to each player than in chess - the sheer number of possibilities quickly becomes overwhelming. To overcome these difficulties, researchers in the 1960s and 1970s began trying to mimic the ways humans use knowledge to reason their way directly to an answer, instead of detouring through all sorts of unlikely possibilities.

Expert Systems: questions and answers

Planes and automobiles
There are only a limited number of factors to juggle when controlling air traffic (left),

but a human brain alone would make mistakes.
AI also takes the guesswork out of tuning a car (below).

The classic example of what expert systems do is provided by medical diagnosis computers, which re-create the reasoning doctors use in diagnosing and prescribing a cure. An early expert system was MYCIN, developed in the late 1960s by Ed Shortliffe at Stanford University, which specialised in diagnosing bacterial infections and prescribing antibiotic cures.

Diagnosing disease involves mastering a collection of fairly simple rules. One rule might be: "If the patient has a fever, consider the following list of diseases..." Another might be: "If the patient has white spots on the throat, consider..." Whereas doctors laboriously learn hundreds of such rules in medical school, a computer can be programmed in a jiffy to apply them.

'When an expert system messes up, it does so big time - they have to be supervised as well as specialised'

Not surprisingly, they do it very well. Since the 1970s, expert systems have on average outperformed doctors in diagnosing most diseases. Computers have faultless memories and iron-clad logic when it comes to combining rules. But not only do expert systems lack bedside manner, they also have important faults that have excluded them from widespread use in medicine.

Machines that tell the human body what's wrong with it
Computers are more accurate than human doctors at diagnosing illness but they need to be constantly fed information - and mistakes can happen.

  • Narrow knowledge Computers easily cope with 50 to 100 rules, but any more confuses them. Just as MYCIN was restricted to bacterial infections, even today expert systems are specialists, useful in a narrow realm only.
    When expert systems mess up, they do so spectacularly, falling off their narrow platform of knowledge into incompetence. This is particularly troubling in medicine: the achievement of an expert system that treated 96 per cent of patients brilliantly would be quickly overshadowed if it killed the remaining four per cent.
  • Communications Given this narrowness, expert systems cannot be left to work unsupervised - and this is a real disadvantage. The constant input needed to keep human and machine working in harmony is too time-consuming, particularly for doctors. This is not true in other realms.
    Expert systems are now routinely deployed throughout banking and industry. They detect credit card fraud and help brokers pick investments for clients. For car-makers they pick up engine faults and administer the complicated rules governing eligibility under warranty. And in customer relations they pass on requests for  help to the right person.
    Expert systems have proved to be good coaches in areas where the difference between a newcomer and an experienced hand is great, but the knowledge mastered with experience can be distilled into a few rules that change little over time. In fact,most users never know they are working with an expert system; they just think new computer is particularly user-friendly.

Neural Networks: from mind to brain

Pattern recognition A net looks for abnormalities in brain waves (above). The same technology identifies planes from any angle (right) - and may soon be able to predict a pilot's next move.

Neural networks are the closest thing to voodoo yet cre ated by the "white magic" of computer science. Data goes in and information comes out, but what happens in between is a bit of a mystery. Also, they "learn".
While expert systems are modelled on the way people think - the human mind - neural networks take as their starting point the human brain. A neural net is composed of hundreds (or thousands) of simple "nodes", roughly analogous to the brain's neurons. These nodes are arranged in layers. Every node has several inputs from the previous layer, and an output to the next layer.
Each of these inputs is an influence rather than a straight- forward yes/no - one input might say "almost certainly", while two others indicate "perhaps not". The node "fires" an output if the sum of the inputs exceeds its particular threshold (which varies between nodes).

'Because it learns in mysterious ways and cannot explain itself a neural net must be taken on trust-or not at all'

The output is combined with other node outputs from the same layer to become the input to the nodes on the next layer. A network learns by adjusting the thresholds of each of its nodes-changing how readily they fire - and by rerouting the interconnections between the nodes and layers.
Suppose you want to make a colour photocopier that can't be used to copy banknotes. You could build a memory containing images of every note in the world, and update it regularly. Or you could train a neural network. As all banknotes have similarities, once a neural network had learnt to recognise enough banknotes it would be able to recognise notes that it has never seen before.

A ghost at the wheel
One of the most impressive things a neural network can do is drive. An army truck developed at Carnegie Mellon University, learnt to copy a human driver - road conditions are input by video and laser rangefinder.

  • The art of net training It has been mathematically shown that neural nets can learn almost anything if given long enough. The problem is how to tell the network what it's doing right, so it can preserve that, and what it's doing wrong, so it can try something else. There is still a fair amount of art involved in this process.
    Neural nets started in the 1980s when David Rumelhart, James McLelland and other researchers decided to try to overcome the limitations of artificial intelligence techniques by modelling computers on the complex interconnections of the human brain. As usual, they succeeded, but only partly.
    Because neural networks learn by experience and not by programmed rules, they bring much more information to bear on a problem than expert systems can and are much better at recognising complex patterns. But whereas an expert system can explain why it has come to a certain conclusion, a neural network must be taken on trust - or not at all.
  • Unknown factors As a result, neural nets can be misleading. In training, a network learned to be 100 per cent accurate in recognising camouflaged tanks. But in the field, it failed completely. Researchers eventually realised that all the training videos with tanks hidden in them had been slightly darker than those without - the vastly complicated neural net had only learned to tell light from dark.

Logical thinking

While neural networkers have tried to re-create the human. brain, and some programmers have created expert systems (see boxes above), others have been exploring the abstract realm of mind. Since Greek times, philosophers have worked to create hard rules of reasoning, called logic. For most of its history, logic's goal has been to discover largely abstract eternal truths through formally constructed arguments.

The advent of artificial intelligence inspired people to try to use formal arguments to describe the messy world around them. Two developments helped greatly in bringing logic into the real world. The first, which began early in the 20th century, was modal logic.
Whereas classical logic assumes a statement is either true or false for all eternity, modal logic can cope with statements that may be true at one time but false at another - for example, "John Major is Prime Minister", which is true now, but won't always be.
The second development, non-monotonic logic, was largely developed by artificial-intelligence researchers in the 1970s and 1980s. Non-monotonic logic addresses the related problem of incomplete information. In day- to-day life, we often make useful assumptions that later turn out to be wrong. "Birds fly" is a good assumption, even if you later discover penguins and roast turkeys.
In modal logic, if an assumption is always false - for example "John Major is a woman" - then the whole system collapses and all its conclusions fall apart. Non-monotonic logic, by contrast, copes gracefully if the bird turns out to be flightless later on. Any conclusions that depend on the erroneous assumption are withdrawn and the rest are still true.

Every time a computer fails in a new job we learn something more about human reason AI and the law AI helps judges deal with petty criminals - freeing courtrooms for important cases

Will computers ever be able to think like a human?
The processors of a computer's chip-brain are made up of hundreds of "logic gates" which only understand 0 or 1. If both inputs (A and B) to a gate are 1, then the output will be 1; if one input is 1 (A or B), the output will also be 1.
This "Boolean" logic - from the 19th-century English theorist George Boole - is powerful. But now computer scientists are developing non-binary alternatives. "Multivalued" logic works with inputs and outputs that take more than two values. Fuzzy logic operates across a spectrum from "no", through "possibly" to "almost certainly" and "yes". The prospect is for computers that think in a less black-and- white way - but that doesn't necessarily mean more like a human being.
Logic that washes Bertrand Russell (left) helped end the rule of abstract classical logic. Now there are many different "logics" - fuzzy logic is used to program advanced washing machines (right)

Thinking like a human

These new thinking tools spurred logicians to try to capture more of the subtleties of real life in their formal statements. This is no mean feat. To reduce humans' instinctive understanding of the world to formal logic, logicians must cope with a variety of philosophical conundrums. Wood, for example, is still wood when it is cut up into bits, but a cut-up table is just so much junk.
What other things are like wood, and how do we distinguish them from things like tables whose identity seems more fragile? Such questions have added fuel to the larger debate about whether the real world - or indeed human intelligence - can be reduced to logical terms. Some argue that the world is fundamentally illogical, and that to reason as humans requires a mass of arbitrary assumptions and leaps of reasoning derived from the experience of living in a body, in time, in the world. To which the logicians reply that although no system of formal reasoning can capture the subtleties of human thought, computers would be no use at all without the guarantee such a formal system provides.

It is out of the see-sawing progress of this argument over computers' abilities that progress in the field of artificial intelligence emerges. Each advance in formalising some aspect of human reasoning (or the world) enables computers to do some new job. And with each new task computers take on, new limitations become apparent.

Fuzzy logic

Even flexible non-monotonic logic treats a statement about the world in terms of either true or false. In practice, however, many things are "sort of" true. Lofti Zadeh, a professor at the University of California at Berkeley, developed a kind of reasoning that uses "fuzzy" terms. Japanese researchers promptly adopted the new techniques to improve controls for washing machines. Instead of trying to define just how dirty is "dirty", fuzzy controls enable a machine to treat, say, an 80 per cent full load of clothes as if they are about 40 per cent dirty.

Case-based reasoning

Humans learn from experience: whatever worked in the past may work in the future. But "case-based" reasoning enables one person to benefit from the experience of others. When a problem is solved, a computer files away the solution and indexes it by a description of the situation. Faced with a new problem, you enter a description and see if anything like it has been tackled before.
Compaq, a maker of personal computers, uses this technique to keep its technical support staff abreast of what can go wrong.

Legal reasoning

Building on research by Imperial College, London, courts in New York and Singapore now use AI in straightforward cases. Not only do the computers speed up justice, they also make intelligent suggestions about rehabilitation.
Each addition to the grab-bag of artificial intelligence techniques broadens the range of tasks computers can perform; it also pushes us towards a fundamental philosophical question. Will all these pieces of intelligence ever add up to the mental equivalent of a human being? Needless to say, philosophers debate the question heavily. The litmus test of philosophical opin ion is a thought experiment pro posed by thinker John Searle. Imagine a man who lives in a tin' enclosed room. He speaks only English, but in the room is a huge Chinese dictionary. Each morning, someone pushes under the door a set of tiles, also written in Chinese. The man's job is to rearrange the incomprehensible symbols according to the patterns given in the dictionary, and then each evening, to pass the tiles back through the door. Is the man's work intelligent?

Some argue that the mans work cannot be considered intelligent because he has no inkling of the meaning of the tiles - and therefore symbol-processing computers cannot be intelligent either. Others reply that humans are overconfident when the argue that they "understand' words and sense impressions and more than a computer does. The debate is a deep and intriguing one, but it may prove irrelevant in the end. From this point of view of the person outside the room, after all, it doesn't really matter if the work going or inside is intelligent or meaningful - so long as it is useful.
John Browning

TED TALK Jeremey Howard Computers that learn
TED TALK Sebastian Thrun What AI is - and isn't.
New Scientist Computer with human-like learning will program itself


Mar95 p30
File Info: Created 14/8/2001 Updated 16/8/2001 Page Address: