Can a machine think? The search for the answer to this question has resulted in recent scientific advances that have important and far-reaching implications for the Navy and the nation. Scientists have produced practical robots which may not think, but which can certainly take over many tasks that until now could only be performed by human beings.
It is the purpose of this paper to summarize the recent progress in the field of robotry so that informed and constructive thought can be brought to bear on its implications. Having served as a forum for introducing new ideas to the naval profession for more than 85 years, the Proceedings seems uniquely suited to acquaint thoughtful naval officers with a matter which may have a greater impact on the human race than any technological development in history.
A robot is any automatic device that performs functions ordinarily ascribed to human beings, or operates with seemingly human intelligence. During the last five years, devices have been demonstrated which can:
Play and win against a group of admirals in a checker game.
Write trivial scripts for TV “Westerns.” Learn to distinguish between various patterns (such as those made by a rook and those made by a knight on a chessboard) without being told which is which.
Read and translate Morse code signals. Compose serious music which is in some respects artistically and technically better than that written by, say, Schonberg.
Invent new proofs to elementary problems in geometry.
Read documents and take action on the basis of what is read.
Spontaneously distinguish music written by Bach from other music.
Read and translate documents from one language to another (but with not very good quality of translation so far).
Predict the plankton content of the sea (something that has been beyond the ability of human scientists).
Distinguish electro-cardiogram patterns of certain heart diseases better than a trained physician can.
Sunday supplement writers often credit such devices with the ability to “think.” Some actually do exhibit properties that bear a superficial resemblance to man’s ability to learn from experience, to utilize intuition, and to improve himself. Thus they can be of great economic use in peace and, conceivably, important weapons in war. Still, once their principles are understood and the mystery of their operation vanishes, no one seriously believes that they really “think.”
It has often been observed that a modern governor on a steam engine, or a household thermostat, would be called a robot by someone living in the 18th century. Today they are commonplace mechanical devices which are not generally thought to operate with “seemingly human intelligence.” Similarly, devices which are truly remarkable robots by today’s standards will undoubtedly be equally commonplace to our children.
It might be said that science is narrowing the gap between the mental abilities of men and machines. This has the double effect of giving us more insight into human thought processes and simultaneously producing machines of greater “mental” power. But the difference between man and machine is very great, and there is little hope of machines equalling human beings for a long time to come, if ever.
What is the field of effort that is producing such things? Experts are by no means in agreement on where robotry begins or ends, or even on a name for the field. “Artificial intelligence,” “neurodynamics,” “bionics,” “intellectronics,” “cybernation,” and “biomimesis” have been used by various groups to describe work of particular interest to them.
Work in pattern recognition, heuristic programming, communications physics, control systems, and mechanical language translation is directed at solving some of the problems in robotry. It is only natural that various authorities, bringing to bear their special interests and background, set down definitions and descriptions of robotry which are confusing to the layman, and which are inconsistent with one another.
For our purposes, a fuzzy but serviceable definition of progress in robotry would be “progress towards achieving mechanical devices which exhibit abilities to solve problems or perform mental functions which have heretofore required human intelligence to handle.” This definition is fuzzy because it does not clearly separate a clever computer program which solves a difficult problem in a “logical” manner from a similar program which solves the problem in, say, an “heuristic” manner.
Actually we know so little about how the brain works that it is uncertain when the mind works “logically” and when it does not. So it is with robots. If they exhibit the property of solving difficult mental problems, we can, from the military or “customer’s” standpoint at least, call them stepping stones along the path towards an ultimate robot, whatever that might be. But we must reserve the right to come up with a new classification when—• and if—we acquire knowledge of thought processes which are adequate to distinguish man from machine.
Robotry as used in this paper excludes effort towards improving the innate mental capacity of natural biological systems.
Simple robots are those which merely incorporate the principle of feedback. The output of a system is monitored by the robot, and the results are utilized directly to vary the input to the system or its mode of operation. The thermostat is an example.
By “directly utilized” we mean that the output signal from the robot corresponds more or less exactly to its input signal; i.e., the input and output signals are simply and highly correlated.
Higher order robots employ mathematical transformations by which the input signal is translated into an output signal in accordance with mathematical rules which can be quite complex. These transformations can be spelled out in advance by the designer and more or less fixed in the robot. Commercial optical readers for handling credit card accounting are examples of such robots.
If, in addition we also incorporate a device which will choose to use particular mathematical transformations more or less by chance, outputs are obtained which cannot be foretold in advance by the designer of the machine. The machine for writing TV Western scripts is a good example of such a device, as are various robot composers of music.
A next higher order robot might have available to it a number of mathematical transformations and a means of learning and remembering which ones best fit various situations. Its output is then dependent on both the incoming monitoring signal and what it has “learned” from previous experience. The checker playing machine which won against a group of admirals, for example.
Somewhat more powerful robots are “Perceptrons” which incorporate a means of developing mathematical transformations which will efficiently classify a wide variety of incoming signals into a relatively few responses. These robots are useful for pattern recognition problems, such as classifying electrocardiogram patterns, distinguishing various musical styles, etc.
Such robots are particularly useful for forecasting when basic relationships are unknown, since they can develop their own mathematical transformations whether or not the designer knows the proper ones to use for relating input signals to output responses. The NAPP or Navy Adaptive Prediction Program is an example of a robot computer program potentially useful for such tasks as forecasting usage of aviation parts, or even of forecasting stock market prices.
In summary, robots developed to date are fundamentally devices which (like human beings) sense the environment and give a response to it. This response can be simple, or it can be so complex that the robot can outperform its human designer in coping with specific environmental situations (e.g., price forecasting or predicting plankton content of the sea). The mechanism is readily understood : mathematical transformations are used to translate input signals into output signals. If the robot has a means of developing its own transformations, and a means of learning and remembering which ones are most appropriate to various situations, the nature of its mechanical construction is still readily understood. But in operation, such a device would seem to be superhuman in the sense that it could solve mental problems beyond the ability of its designer.
So far, robots have been described which have already been built and are working today. How will future robots work?
A principal objective, I believe, should be to extend the variety of mathematical transformations which the robot uses in translating input to output signals. The human being uses an astonishing variety of such transformations. We can, for example, count objects and readily distinguish 24 from 25 objects— no matter what they are. We can immediately recognize a cat no matter from what angle viewed. We can with facility scale up or down and recognize a model airplane as equivalent to a jet transport. Robots cannot now do these things easily.
Artificial neurons and electrical circuits incorporating them (called “neural nets” by workers in the field) may eventually provide the mechanism for generating the wide varieties of mathematical transformations that robots will ultimately need. Alternatively, large “libraries” of mathematical transformations can be deliberately built into the robot for use in particular situations. In any event, it is clear that more powerful robots require a wider variety of mathematical transformations than current robots utilize.
So far, the term “mathematical transformation” has been used in a limited sense: the output signal is a function of one or more variables. The next task will be to construct robots whose output signals are functions of one or more functions. It is not clear how such robots should or could be built. Possibly neural networks can be devised whose generality in obtaining transformations is far beyond anything yet accomplished in biological organisms. If not, some combination of neural networks with more prosaic function generators and/or analyzers is indicated.
The second direction of future effort on robots will be to increase the self-analysis of the robot device itself. Self-correcting and even self-healing (or repairing) responses are an obvious corollary to input signals derived from the robot sensing itself as part of the environment. Developments in this direction are of particular interest to the military services since they may usher in devices of far greater reliability than anything yet known. In the years to come, it would not be unreasonable to consider cruisers manned by a dozen human beings or large supply centers with scarcely a person in sight.
The combination of greater versatility and greater reliability in robots of the future will clearly pave the way for new techniques and for new devices of military importance.
In a recent article,* Lieutenant Commander John Chastain, U. S. Navy, pointed out that the NTDS (Naval Tactical Data System) as presently programmed is potentially an electronic straightjacket for the commander—a potential Frankenstein who will usurp authority without taking responsibility for the results.
Here is one area where certain robot concepts will ultimately help alleviate a problem of substantial naval significance. The basic problem is how to evaluate and act upon a great deal of information coming into a tactical command center. The way the Navy is having to do this at present is to try to anticipate all possible combat situations which might arise and to trigger off a series of packaged, pre-planned, rigid responses via the NTDS when such situations are detected. Depending on the experience and skill of the computer programmer, these responses may—or may not—be the best ones.
Progress in robotry might assist in solving this problem in three ways. First, the underlying theory might be exploited to give us a more intelligent coupling of man and machine—a clearer understanding of the proper role of each in the total system. It is possible that some NTDS programming had to be based on the assumption that the machine should take over the maximum possible duties rather than the optimum set of duties.
Second, robot systems might be utilized in advance of future NTDS programming to develop both concepts and programs on a laboratory basis. Simulation techniques have made rapid strides in the past two years, and very sophisticated simulations of combat situations could be developed in the laboratory and analyzed with the aid of both conventional and heuristic computer programs. These simulations could be particularly fruitful if used to study various feedback techniques and the effect of internal organization on the end-product of the command control system as a whole. Work in this general area of human decision making has been started by the Air Force at the Electronic Systems Division, Hanscom Field.
Third, there is no reason why future NTDS programs should not be developed to incorporate self-improving features. To use an elementary example: the NTDS program for replenishment of the ship is presently based on filling up storerooms to a fixed level of material. There is no reason why this level should not be modified on a continuous basis by feeding back to the NTDS information as to the effect on the ship of actions taken in response to the program. Computers cease to be electronic straightjackets when they are programmed to monitor the effects of their recommendations and to change the recommendations continually to meet the needs of their human masters. Determination of goals is the proper province of the commander; automation of certain subsidiary decisions as to how best to reach these goals is the proper province of robot equipment.
Military uses of robots are by no means confined to the design of combat decision systems. The following paragraphs give only a few examples of applications which the future may hold.
Detection devices offer one of the most fruitful fields for robots. Sensing gear of many varieties are able to convert environmental stimuli into a series of electrical pulses. The patterns formed by these pulses may be directly analyzed by robot equipment well within the current technological envelope, and the presence or absence of various targets detected. Such equipment is likely to be more accurate and more reliable than current radar and sonar systems which must first convert the electrical pulses into oscilloscope patterns or audible signals for analysis by human beings who are often confused and distracted by clutter or noise.
A similar problem is that of mechanically analyzing seismograph records or aerial photographs to detect presence or absence of features of interest. Reliable devices to perform this function may be available within a comparatively few years.
Substantial numbers of military and civilian personnel are now required by the Navy to perform purely clerical functions such as keypunching; sorting, matching and routing documents; transcribing drawn or written material; copying Morse code signals; translating material into or from foreign languages; and so on. A minimum of several tens of thousands of people could be used for more important tasks if these functions were automated. Prototype robots to do all of these things have been demonstrated. Further improvement in reliability and reduction in cost will lead to widespread use of these machines in the Navy during the next decade.
Laboratory uses of robot equipment are important. Monitoring of cloud chamber traces for detection of nuclear events and analysis of astronomical photographs are fairly obvious uses for pattern recognition devices now in the prototype stage. Somewhat further in the future is the coupling of man and machine for development of fundamental knowledge—even such abstract knowledge as mathematical theory. The literature of science, and particularly of mathematics has long since passed the point where any one human being can have even reasonable familiarity with all of it.
Robots could be devised to store at least the written literature, rapidly search it for promising leads, and to arrive at tentative theories and solutions on the basis of abstract notions of “goodness” of the end product. By optimally coupling a skilled scientist with a well designed robot, it is possible that stupendous breakthroughs will be made in mathematics, medicine, and physics to mention three of the more promising fields.
Process control functions, although not as widespread in the military services as private industry, represent an easy application for robots. The basic problem is to monitor a number of operations contributing to a set of goals, and to readjust the operations as necessary to meet the goals more efficiently, or to conform to a new set of goals. Middle management decisions to do these things, which are now made by human beings, can be readily automated with heuristic programs on conventional computers. It is also feasible to control assembly line operations so that more or less “custom made” products with numerous individual differences (such as electric meters) can be made on a mass production basis.
From the foregoing examples, it is clear that robot applications will move steadily upward from routine clerical operations (such as reading documents) through middle management decisions optimizing paths towards goals (such as process control) to top management decisions involving balancing among several goals (such as command control center equipment). Is there a limit? Can we hope that robots can ultimately set goals? Can they ultimately devise such apparently new ideas as the theory of relativity or the calculus or the social concept of church- sanctified marriage or the unified philosophy of Buddha?
There are many who doubt it. A. M. Turing has listed nine objections to the idea that a machine can attain the power of human thought. These objections range from theological reasons to arguments based on physical differences (e.g., machines have no emotions). In my own opinion, there are at least two real limitations on machines as competitors with human beings in a thinking contest. The first is that there is some evidence of extra-sensory abilities in human beings. Unless machines develop these abilities as they grow to the size of the human brain, there may be times when a human being can use psychic phenomena to best a machine.
The second limitation may be temporary, but it is surely with us now. We know from Goedel’s theorem that no discrete state machine can answer all questions which can be posed under a self-consistent system of logic. So far as we know, however, human beings do answer such questions (perhaps by shifting to another system of logic) and such answers appear to solve satisfactorily the real problems that human beings face. Robot devices so far proposed are discrete state machines operating within the framework of a single system of logic. It may be possible to devise equipment that will get around this objection, but so far it has not been done.
I think that within the next few decades we can look forward to the use of robot equipment for making increasingly higher level decisions, including those involving fundamental national strategy, social goals, and political questions. Certainly there are limitations on machines. It is probable that some sort of chain reaction will be set off in a decade or two whereby a machine will be built which will, in turn, be able to build another machine that is, in some sense, better than itself. But there still appears to be the unique human ability to use extrasensory phenomena, and to operate freely within a number of not necessarily consistent systems of logic, which will forever doom the machine to a subsidiary role in conceiving and selecting goals.
Meantime, it is clear from the very few examples outlined above that robots will affect practically every aspect of military life, and indeed of all human endeavor. Constructive thought and discussion should be encouraged so that important applications will not be overlooked in the widespread race to utilize artificial intelligence equipment for specific purposes. These are all too frequently related solely to whims of scientists active in this new technology, rather than to over-all national goals.
Both the general public and scientific decision makers at the national level have given greater attention to robotry in Russia than in the United States. As far back as 1956, A. N. Nesmianov, then President of the Soviet Academy of Sciences, stated that understanding the nature of the brain and the modelling thereof (i.e., construction of artificial intelligence devices) was the most important task facing Soviet science. The basic concepts of cybernetics are widely known to the Russian public, and a great deal of attention has been given it in connection with the drive towards automation.
The average Russian seems to be more excited by the prospect of using robots to relieve him of his daily tasks than is the average American. This may be due to greater attention given cybernetics by the scientific and engineering community as a whole, and the consequent carry-over of this interest to the popular press and to statements of political and social leaders.
Observation of Russian progress as revealed by both the literature and personal contact have led to the general opinion that the Soviets are well aware of the potential of robots; they have devoted substantial resources to them; and they are competent. Of considerable interest is the Laboratory for Electromodelling (of the brain) headed by Professor Gutnmakher in Moscow. Gutnmakher’s published papers display a keen intellect, but have not discussed his ideas on solving the fundamental problems of constructing really worthwhile brain devices. It is of interest to note that he now has at least several hundred scientists from many disciplines working under him. No organization of comparable size working exclusively on artificial intelligence problems exists in the United States.
Robots in general and artificial intelligence in particular offer a greater military (and human) payoff than atomic energy, both within the next decade and in the long range future. The Russians seem to be more alert to the potential, and more serious about exploiting it than we are. Is an all-out national program similar to the atom bomb’s Manhattan Project of World War II called for?
This problem is simply answered. There are probably less than two hundred scientists in the United States making really useful contributions to the field of artificial intelligence today, and some of these would not work well under even the most skillfully directed project organized to achieve end products in the form of devices to perform specified functions. Furthermore, diversion of the effort of the remainder towards practical hardware would lessen the possibility of achieving any real breakthrough. The present state of theoretical knowledge is simply not adequate to sustain an engineering effort of the magnitude implied.
For these and other reasons, I believe that a program similar to the Manhattan Project is not the answer. Nevertheless, we face a problem of national danger, and perhaps of national survival. More people, more money, and better co-ordination of effort among the many academic and industrial groups working on various robot problems would probably produce fertile ground for really significant breakthroughs, and would certainly produce low-level but useful devices at an earlier date. For example, I suspect the problem of building an optical reader able to read anything a human being can read, and do it more economically, could be solved within the framework of existing theory and technology by a few dozen competent men in a matter of a few years.
The problem of mechanizing language translation to give a high quality output would require more people, and a matter of many years. Speech typewriters, automatic chauffeurs, and other devices are further over the horizon. But a breakthrough is needed before a robot can be built that will look around the room and decide what is trash and what is not. And a real breakthrough is needed before problems of national strategy and social goals can be attacked in some meaningful fashion by machines.
It is this fundamental necessity to provide the climate for breakthroughs such as are needed that suggests a national effort to do these things:
(1) Attract additional competent scientists and engineers to work on problems of artificial intelligence—particularly people from the biological disciplines who have a great deal to offer.
(2) Provide more—much more—resources in the form of money and facilities; and develop the proper balance between freedom to pursue individual inclinations and co-ordinated effort towards common goals.
(3) Improve dissemination of scientific knowledge in the field of artificial intelligence. Automating the flow of written information is part of the problem; increasing and making more efficient the amount of personal contact among scientists with related interests is another.
I suggest the above goals can best be accomplished by assigning the Office of Naval Research the task, and giving this office the funds and the freedom to do it. To organize the effort under another agency of the Government would involve starting from a less experienced and perhaps less competent base. The Government should capitalize on the hard-won respect of the scientific community which ONR has gained over the years; and it should build on the nucleus of able personnel in this organization to put together a national effort which will guarantee our pre-eminence in the development of artificial intelligence—a major key to man’s further progress and perhaps his survival.
Meanwhile there is a real need for naval officers to become familiar with the potentials of this new field. Out of informed discussion and widespread understanding of these potentials, it is probable that some sort of priority list of end uses can be developed so that scientific effort can be properly guided during the critical years just ahead. More importantly, the naval profession can begin the task of adapting its doctrine and future plans to use this new tool of science to the best possible advantage in furtherance of this country’s national goals.
* See John Chastain, “The Role of Computers in Combat Control,” U. S. Naval Institute Proceedings, September 1961, p. 59.