Computer intelligence: unlimited and untapped. (an interview with Herbert Simon) Betsy Staples.

Computer Intelligence: Unlimited and Untapped

"It seems fairly clear to me that there are no discernible limits to the range of things that computers can be programmed to do.' The speaker is Herbert Simon, Nobel prize winning economist and professor of computer science and psychology at Carnegie-Mellon University.

For nearly half a century, Simon has been involved in the study of human decision-making and problem-solving processes. And for the past quarter century he has been using computers as tools for both the simulation of human thinking and the augmentation of thinking with artificial intelligence.

The venerable professor, long a favorite of CMU students, shared some of his views of the future and his perspective on the history of artificial intelligence at a recent meeting of the Carnegie-Mellon Business Club in New York City.

In recalling that "back in 1955, a few of us, including Ai Newell and myself, decided that there were some more interesting things you could do with computers than simply have them make the payroll,' Simon reminded his audience that "computers aren't just number crunchers, even though, unfortunately, most of them are still condemned to spend their days crunching numbers.'

He pointed out that in order to make computers effective number crunchers, they had to be given the much more general ability to operate on symbols, including words.

"We have all begun to hear about word processors and to realize that the English language is just as congenial to computer memory as algebra or arithmetic,' and "once you have given a system the ability to process symbols, you have given it the ability, with appropriate programming, to think--to do those things that a human being does when we say that he is thinking.'

To discover whether a human being is thinking, Simon said he would give him a problem to solve. He thinks we can apply the same test to a computer: "So the whole field of artificial intelligence (about 27 years) is directed at inducing computers-- programming computers--to do the kinds of clever things that human beings are capable of doing when they think.'

There are two objects to this "game,' he said. The first is to increase the productivity of computers--that "vital and productive resource in our society'--in new areas. We should not be willing to limit this increase in productivity to business and scientific computation, "but to bring the computer in as an augmentation to human thinking in all the domains in which human beings think.'

"The second half of the venture,' he said, "has been to use the computer to simulate human thinking to get a deeper understanding of how we, as human beings, think--what the processes are.' He illustrated his point with a discussion of computer chess programs.

"Computer chess programs on the market demonstrably do not play chess like human beings. The very good and most powerful ones typically look at a half a million to a million possibilities before they make a move--the game tree: if he does that, I do this, he does that, and so on. Even your little home computer looks at tens of thousands of branches on that game tree.

"No human being does. We have evidence that a human being in a difficult position, expert or amateur--even pondering a chess board for 15 or 20 minutes --probably doesn't look at more than 100 branches of the game tree. The difference between an expert and an amateur is that the Grand Master looks at only the right branches.'

Expert Systems

Other programs that make use of the ability of the computer to store large amounts of information and draw conclusions from it are what Simon calls "expert systems.' A program called Caduceus, for example, is "a pretty good diagnostician in internal medicine--good enough that if you are a physician and you have a difficult case, if would probably be worthwhile to bring the program in as a consultant--if only to see if it has some different ideas of what's wrong than you have.'

How does such a program work? "Well, first of all, it has a tremendous amount of medical knowledge gleaned from text-books and clinicians--a tremendous medical data bank.' But the data bank can't perform the diagnosis, so the program must be able to draw conclusions based on the knowledge in the bank.

"In fact,' said Simon, "if you look at the structure of Caduceus, you see that the kinds of reasoning it does are similar to the kinds of reasoning a human diagnostician does. It forms hypotheses (it could be this or could be that); it asks for tests to be made that will help it discriminate among the hypotheses; it begins to weigh the evidence and rule out certain things; and finally it arrives at a diagnosis for the case.

"We see a great burgeoning of these expert systems, and I don't care whether you want to call them intelligent or not; the plain fact is that they make use of information in arriving at professional decisions at the level that a good human professional in the area in question can arrive at those decisions.'

He added that in some areas such programs have become so good that they are of great use to professionals in their fields. Chemical manufacturing is one area in which there are programs "that do a very sophisticated job if you have a certain organic molecule you want manufactured. They do a very sophisticated job of finding good reaction paths to enable you to do that, taking into account the costs of raw materials, the thermodynamics of the reactions, and what not.

"To the best of my knowledge, such programs are now in regular use, partly in automatic mode, partly in interactive mode with chemists--you can use these ideas in both ways.'

State Of The Art

In discussing the state of the art in artificial intelligence, Simon commented that "it is really only within the last five or ten years that computing power has become cheap enough and powerful enough to make many of these schemes cost effective as distinguished from intellectually interesting.' The advent of microcomputers has moved artificial intelligence into the realm of practical applications, he thinks.

Most people, he said, learn that computers are tools that are limited to quantitative mathematical modeling based on theories of optimization--linear programming and queuing theory, for example-- and that it is very easy to find real world situations that are too complex for these tools to handle--"either because the situations contain qualitative elements or because the mathematics gets too hairy. So you have to resort to common sense.

"Or you take a problem, squeeze it until it appears to fit a mold, and then solve that problem and hope it has some relevance for the real world.' He cited the example of scheduling in a job shop. You have a theory that works only when all the orders arrive simultaneously and are scheduled, and then nothing else ever happens again. Would that we could have a job shop like that!

"What is happening now as part of the Robotics Project at Robotics Institute at Carnegie-Mellon is that people are trying to see what you can do with job shop scheduling if you apply the techniques of artificial intelligenc--that is, if you make it not a task of solving a mathematical model, but a task of reasoning, of making inferences about a large and changing database.

"The large and changing database is that information that describes both the orders as they come in, the shop capabilities, your plans, inventories, and so on. So you are solving not just an optimization problem; you are applying rules of thumb or heuristics in your attempt to keep the schedule going as the situation changes.' Such a system is now operating on a test basis for a large manufacturer of turbine blades.

Robotics

"If there is anything revolutionary about roboticc,' he said, "the revolution is still to come. Robotics will become revolutionary when these devices are so flexible in their sensory capabilities, so flexible in their effector capabilities that you don't have to very carefully shape and smooth the environment in which they work.'

He pointed out that in the past, mechanization has been more dependent on changing the environment in which the mechanized devices worked than on building the devices. For example, to make use of a given mechanized device, you might have to ensure that the floor in the area in which it was to be used was smooth enough to allow it to roll around.

If the floor could not be made smooth enough, you would not be able to use the device.

"What we are looking for now,' he explained, "are advances in the art of building receptors--ensory organs--for machines which are clever enough to extract patterns--for example, looking around a room and seeing that there are people in it, or even being able to count them. That, I think, is still beyond--or at--the frontier of the state of the art in robotics.

"If you are worried rather than challenged by the prospects of lots of robots around the world, don't hold your breath. The rate of progress in robotics will be dictated primarily by the solution of these pattern recognition problems . . . and secondarily by the development of more flexible effectors--particularly effectors that have high strength to weight ratios, so that every time they pick up an egg they don't break it.'

Another problem that Simon sees central in robotics today is that of reducing inertia in machinery. He thinks "it will be solved primarily by mechanical engineers and specialists in materials rather than by AI specialists.

"What I am saying is that there are some very fundamental problems of applied research that must be solved as we try to extend robots into new domains. In some ways, at least, the whole picture has been oversold a little bit--for the short run, not necessarily for the long run. The rate of change, I think, has been exaggerated a little bit.'

When he talks about the future "and the range of activities over which we might expect to see computers doing something interesting,' he is haunted by the question "Isn't it true that computers can do only what you program them to do?'

"Literally, of course, that is true. You can write a program that is capable of undertaking a search or a program that is capable of learning; however, there is no reason to suppose that when you wrote the program you knew what the computer was going to learn, or what problems it was going to solve, or how it was going to solve them, or even that you could solve them yourself.

"By the same token, the fact that you or I wrote the program does not mean that the computer cannot do anything creative --cannot discover anything which we couldn't discover or which hadn't been discovered by human beings.'

Bacon And Ohm's

As an example of a program that can discover things, Simon cited a computer program called Bacon (for Sir Francis, of course). He calls it "an inductive machine.' He went on to explain: "You give Bacon data--raw data--and Bacon's task is to find the scientific laws that are hidden in the data. We have tested Bacon primarily on historical scientific discoveries. We said to Bacon, "well, if you think you're so smart, let's see what you can do by taking the data that Kepler had about distances of planets from the sun and their periods around the sun.' Bacon discovered Kepler's Third Law in about 59 seconds. It also discovered Ohm's law very rapidly-- not by trying all possible laws, but by following a few selective heuristics which led it to look at plausible things first and less plausible things afterward.

"If, in the course of developing such laws, Bacon discovers that there is some kind of lawful relation between several objects, Bacon will try, in order to arrive at laws, to introduce new concepts, new properties of those objects. And so Bacon has re-invented the concept of inertial mass.

"It was given some data that involved the mutual acceleration of two bodies, and it found that there was always a constant relation for a pair of bodies between how fast one accelerated and how fast the other accelerated. We all remember conservation of momentum, I'm sure.' Note that Bacon was only given the data on acceleration. It had to invent and introduce a new concept, the mass, to explain the observed data on the accelerations.

"Given data about the refraction of light, Bacon invents the index of refraction; given data about the mixing of hot substances in the equilibrium of temperature, it invents specific heat. So, it is a program for inventing new concepts.'

Where Do We Go From Here?

In discussing the questions "Where does this go?' and "Where does this stop?' Simon offered the opinion with which this article begins. There are, he feels, no limits to what can be done. He thinks we should ask ourselves: What do we want computers to do in the world? How do we want them to augment our own powers? What do we want them to do just for fun? Do we want them to explore the world of ideas?

He expects computers "to extend their range of use and application much more rapidly in the area of human white collar and executive work than in the area of blue collar work. Both are going to expand, but expansion in the former will be much more rapid. One of the big lessons of the 27-year history of artificial intelligence has been that it is much easier to automate a college professor or a businessman than it is to automate a bulldozer operator.

"The boundaries are moving faster with respect to computrs doing those kinds of things done inside the central nervous system--the thinking kinds of things, the problem-solving things, the use of infor-mation banks in relation to intelligence-- than with respect to physical robotics.'

Simon discussed several aspects of robotics and then threw the presentation open to questions from the audience. These covered a wide range of topics from the Japanese research effort (no threat in Simon's opinion) to the Turing test (he feels that Ken Colby's simulated paranoid program essentially passed the test) to the emotional content of thinking (go slow if your decision-making has high emotional content). However, Simon's answer to one question pretty much sums up his views about computers and AI.

Question: Are there any criteria today that we could use to conclude that computers can't think?

Answer: "I can't think of any; maybe a computer could. I really find no reason to regard myself as thinking in ways different than a computer can think. Fortunately it doesn't bother me, so I don't lose any sleep over it.'