Scientific and Technological Perspectives

COMPUTER systems featuring artificial intelligence (AI) technology may finally provide relief for overburdened commanders. The popular press has been replete with stories about the unlimited horizons of AI applications. Military, academic, and commercial research facilities are investing significant sums to exploit this new technology. But can AI technology really help military commanders? How much of the media coverage is simply "hype" without substance?

The Defense Advanced Research Projects Agency (DARPA) recently launched a $600 million Strategic Computing Program that could lead to completely autonomous weapons, battlefield management systems, vision and speech systems, and an automated copilot that can understand a human voice. Near-term applications are to be in the areas of tactical targeting and natural language interfaces. General Robert T. Marsh (former Commander, Air Force Systems Command), in an article previewing future technology, discussed the potential of AI:

Realizing AI's full potential depends not only on technological developments but also on maintaining an awareness of how these new techniques can be intelligently applied. We must avoid being rushed into unsound AI projects. Computer technology already has a tremendous impact on all facets of our daily lives. We are now entering an era in which computer systems may come to dominate the central core of our existence. Fifth-generation computers, for example, may provide legal and health advice, control transportation systems and traffic flow, educate our young, and serve as lifetime personal advisors. Knowledge and information may become the critical commodities of power in the future. To survive in this new environment, military leaders will have to become aware of the promises and associated problems of artificial intelligence technology.

What Is Artificial Intelligence?

The problem with modern communications systems is not that they cannot provide support, but rather they provide an overabundance of data. The difficulty, in fact, is that commanders have too much information. We seem far better at producing systems that churn out data than at developing machines that sort out the superfluous. When computers were introduced into the military, it was hoped that they would help limit data to meaningful information. Unfortunately, the opposite has been the case. The amount of data that commanders must sift through has increased as computer support systems have multiplied. Now, yet another promise of relief is being discussed. Artificial intelligence is being hailed as the long-awaited computer breakthrough that will provide effective decision support systems for the future. Will artificial intelligence computer systems actually help make a commander's job easier, or will they merely add to the burden? to answer this question, we should look first at what the term artificial intelligence means.

Artificial intelligence does not refer to facts or information about a potential adversary. Rather, intelligence in this context refers to the power or act of understanding. There is no doubt that AI means different things to different people. To the nontechnician, it could mean mystique; to researchers, a specific discipline with complex problems to be solved. The Handbook of Artificial Intelligence (1981), edited by Avron Barr, Paul Cohen, and Edward Feigenbaum, describes artificial intelligence as "that part of computer science concerned with designing intelligent computer systems, that is, computer systems that exhibit the characteristics we associate with intelligence in human behavior––understanding language, learning, reasoning, solving problems, and so on."² In other words, artificial intelligence is an attempt to give machines the capability of performing intelligent human-like tasks. The recent media hype belies the fact that men and women have been working for years to achieve this goal.

Pamela McCorduck in Machines Who Think (1979), perhaps the best history of artificial intelligence, discusses man's continuing attempts to replicate his own abilities. Modern-era efforts to create artificial intelligence began with the advent of the first electronic "calculating machine," ENIAC, in 1946. Pioneers in the field of AI, such as Herbert Simon and Allan Newell of Carnegie-Mellon University, Marvin Minsky of the Massachusetts Institute of Technology, and John McCarthy of Stanford, realized that machines could be made to manipulate symbols for words or thoughts.³ The mainstream of computer development during the early years dealt with straightforward numerical or data manipulation. A few individuals struggled with the concept of creating machines that could demonstrate reasoning and learning capabilities. Early projects centered on games, such as checkers and chess. If a machine could be made to play chess reasonably well, it was argued, then machines could be considered intelligent.

By the late 1950s and early 1960s, articles on computers were already discussing both the promise of this new technology and its threat to man's superiority. A 1961 Life magazine article titled "The Machines Are Taking Over" stated that computers were slowly replacing man in many endeavors but that man could always "reach down and pull the plug."⁴ Sales executives at IBM were afraid that their computers would be psychologically threatening and customers would refuse to buy them. Ads were developed to show that computers were really pretty dumb after all.⁵ Arguments over whether machines could actually think were initiated with these early computer developments.

Can machines think? No present models for understanding knowledge formation or how the mind works allow us even to begin to answer this question. Certainly this article, as an introduction to AI, will not dwell on what may be a moot point. But we can examine a few of the basic arguments surrounding this debate, and perhaps by so doing, we can remove some of the mystique surrounding AI and provide a common basis for understanding such terms as intelligent machine or smart systems.

Early researchers who said that machines could be made to demonstrate human-like intelligence developed the "Turing Test" (named for A.M. Turing, a British AI pioneer) to prove their point. In the test, an interrogator would be separated from the person or machine being interrogated and could communicate only by teletype. If the interrogator could not tell whether he or she was communicating with a person or machine, then a machine could be said to think.⁶

Opposite these believers were those who questioned the very idea that machines could be intelligent. The following is a summary of the primary arguments against the concept of thinking machines:

Despite such arguments, a number of AI researchers today believe that some machines do perform thinking functions. They argue that just because computers can't write like Shakespeare does not mean that they aren't intelligent. AI expert Patrick Winston stated the case in this way: "Of course to believe in human superiority is a tradition. Once our intelligence was unchallenged, yet someday computers may laugh at us and wonder if biological information processors could be really smart."⁸ The arguments go on and on. Regardless of the position taken, it is a fact that AI developments will require computer systems that are physically and operationally different from conventional computers.

To understand how AI systems work, first let us briefly review a few fundamentals about conventional computers. Computers, in general are devices that accept and manipulate data in, a sequence ordered by some prearranged program. These operations result in some further action or output. Computers that perform these operations are generally divided into two basic types-analog and digital. Analog computers operate on a constant but varying input (like an automobile speedometer), while digital computers operate on inputs that are either on-off or incrementally stepped quantities represented by numerical digits.⁹ AI systems employ digital computers.

Digital computers have three main components: an input/output device, a memory module, and a central processing unit (CPU). The input/output device (keyboard, monitor, printer, etc.) provides the means to enter programs and to display or view results. Programs and instructions are stored in the second basic component, the memory module. Interim results, computations, and data are also stored in memory until they are needed for further operations. Memory modules may also use storage devices such as magnetic tape or discs. The key component of a conventional computer, the central processing unit or CPU, processes the programs or instructions in the memory module and executes the required operations. It controls the entire operation.¹⁰

All conventional computers, from the first generation machines built in the late 1940s and early 1950s through the current fourth-generation systems, are essentially the same in design and operation. Generational dividing lines came about as a result of changes in hardware technology rather than operational techniques. First-generation machines, for example, used vacuum tubes, created a great deal of heat, and were very large. Second-generation machines featured transistors that reduced both size and heat problems. Integrated circuit computers introduced the third generation, and very large, scale integrated (VLSI) computers initiated yet a fourth generation. Edward Feigenbaum, a leading AI expert, believes that we are currently at the end of the third generation and that fourth-generation VLSI (computers) will dominate the 1980s.¹¹

Conventional computers built during all four generations follow an operational design known as the Von Neumann process. (John Von Neumann was a computer pioneer and mathematician.) This means that computer programs are processed serially in a step-by-step operation. Each step that the computer takes is spelled out in a detailed program. It can do only what it is instructed to do. It cannot assimilate new facts that were not included in the program, and it cannot be creative. A conventional computer is simply an arithmetic machine that receives data, performs simple arithmetic, and produces answers consisting of individual digits. Special programs in the computer can convert individual digits to alphabetic characters.¹² Conventional computers, then, follow rigidly formatted programs, completing one process at a time; but technological improvements have enabled conventional computers to perform these tasks at remarkable speeds. Artificial intelligence computers operate in a fundamentally different fashion.

Artificial intelligence systems differ in both their hardware and operational programs. AI computers are built to manipulate symbols rather than numeric values. These special computers are made primarily by three companies: Symbolics of Cambridge, Massachusetts; Lisp Machines of Culver City, California; and Xerox Electro Optical Systems, Pasadena, California. These machines are constructed to use unique AI programming languages such as LISP (List Processing Language). LISP was developed by John McCarthy in 1957 for the express purpose of handling complex concepts and symbol manipulation.

Conventional computers and AI systems also have a number of significant differences in the way they operate. You will recall that conventional systems use primarily numeric operations, following very precise step-by-step directions. That is, to solve problems, they follow explicit algorithmic solutions. Data and operational instructions are part of the same program. Because information and instructions are structured, it can be very difficult to modify or change a program. Conventional computer programs are designed to provide specific answers to a given problem. They are not designed to guess, but rather to process data and provide solutions stored in the computer's memory. It is this inflexibility that led AI researchers to design machines that could simulate more flexible human thought processes.

Artificial intelligence systems are primarily symbolic processors. Rather than following a predefined algorithm, the AI program sorts through its stored memory to determine its own sequence of steps. In this approach to problem-solving, solution steps are implied but are not specifically spelled out. The ability of AI systems to use "heuristics," instead of merely preset algorithms, gives them their most unique characteristic. Heuristics have been called the "art of good guessing. " Heuristics enable us (or machines) to recognize promising approaches to solving problems, to break problems down into smaller problems, to overcome incomplete information, and to make educated guesses.¹³ It is this flexibility that enables AI systems to develop satisfactory answers that may not be precisely correct but are acceptable. Another important aspect of this flexibility is the AI system's ability to explain why certain decisions were made. In an AI system, the knowledge base is separated from the instructions on what to do with that knowledge. As a result, programs can be modified easily, or new data can be added to the knowledge base. Knowledge "engineers," new technical specialists, have the job of capturing and translating expert knowledge into AI data bases.¹⁴ Table I provides a comparison of conventional and AI systems.

Artificial Intelligence	Conventional Computer

Primarily symbolic processes	Often primarily numeric
Heuristic search (solution steps implicit)	Algorithmic (solution steps explicit)
Control structure usually separate from domain knowledge	Information and control integrated together
Usually easy to modify, update, and enlarge	Difficult to modify
Some incorrect answers often tolerable	Correct answers required
Satisfactory answers usually acceptable	Best possible solution usually sought

Thus, artificial intelligence is not a new field of endeavor, but it does use computer technology that differs from that of the conventional computing machine. The major challenge facing developers has been to find ways to apply AI systems effectively.

AI Applications

Converting AI into practical applications has not been easy. During the 1950s, for example, enthusiasts voiced extraordinary claims for this new technology. DARPA funded a computer program to translate Soviet documents into English. The difficulties of AI machine translation became clear when the Russian term hydraulic ram was translated as "water goat." Despite such setbacks, DARPA continued to almost single-handedly keep AI research alive in the United States. During the past two decades, DARPA has invested more than half a billion dollars in various types of computer research.¹⁵ Because of this continuing support, equipment is now available to develop practical AI applications.

Another major investor in the future of AI is the Rome Air Development Center (RADC), which is spending more than $7 million per year on AI research. Application areas being studied include speech processing, tactical mission planning, intelligence data analysis, and software development.

Simultaneously, commercial companies are trying to apply AI technology, using expert systems for tasks ranging from diagnosing medical problems to helping repair cable systems and diesel locomotives. Expert systems are also helping to discover new oil and mineral deposits. Business Week reported that "optimistic analysts are predicting that AI will become a multibillion-dollar annual business well within a decade."¹⁶

Both military and commercial researchers have looked at the possibilities of applying AI technologies to vision systems. DARPA and the U.S. Army Engineer Topographic Laboratories (USAETL) have been trying to develop systems for years that could interpret imagery automatically. Finding a system that can learn to differentiate among various patterns and objects may be one of the toughest challenges AI researchers face. To make it easier to use these and other AI systems, some AI researchers are trying to develop "natural language" systems.

Natural language systems offer hope for all those who would like to use a computer but have neither the time nor the inclination to learn formal computer languages. Such a system would allow an operator to talk naturally to the system. The burden of understanding would be on the machine rather than on the human. Intellect, a commercial software program, converts typed natural language (English) instructions into machine language. It then translates the instructions back into English and displays them on a monitor so the user can confirm that they were understood.

Automatic speech recognition (ASR), another form of AI natural language, is also being investigated. RADC has worked for more than ten years to develop systems capable of automatically interpreting speech, picking out key phrases, and identifying the speaker.¹⁷ The problem is extremely complex. Humans interpret speech in the context in which it is heard. Even when words are run together, humans can pick out the ends of words, phrases, or sentences. Computers cannot yet understand continuous speech by a random speaker.¹⁸ In natural language research, as in other areas of AI, a number of problems must be solved.

Given all the progress to date, one must understand that there are still many problems with AI technology that have not been solved. The difficulties range from misinformation and consumer confusion to specific technical difficulties. A number of AI publications now available offer lengthy discussions about technical developmental problems. Only a few such problems will be addressed in this overview.

Although AI has been researched for almost three decades, the number of AI experts is very limited. For example, the few knowledge engineers available are converting knowledge bases into machine coding largely by hand and are likely to continue to do so into the foreseeable future.¹⁹

Generally, military computer programs are written in a rigidly formatted language, such as FORTRAN. Complicating the picture even more may be the fact that the Department of Defense has adopted ADA (another highly structured language) as the official program language for embedded computer systems (missile guidance, for example). However, AI computers must use a more flexible language, such as LISP or PROLOG. If AI is to be generally accepted for military application, the computer language compatibility problem will have to be solved.

Many of the difficulties associated with AI are being downplayed by enthusiasts, while at the same time AI is being oversold. Some AI researchers are afraid that the media hype may have created expectations that cannot be met. They are concerned that there will be a backlash similar to the one that followed the disastrous failure of the machine translation effort in the 1950s. The layman's difficulty will be to separate facts from overzealous promises. Business Week reported in July 1984:

Computer technology has developed at an incredible pace. The world is transitioning into a society that lives on information. Traditionally in the past, national power has been measured by such elements as territory controlled, annual production output, military troop strength and arsenals, and so forth. A new basic element of power may be added to the list. In the future, nations that control information or knowledge may possess a major source of influence in international affairs. The systems that make such control possible are likely to be the products of the so-called fifth generation of computer technology. These new systems will represent a distinct break with conventional Von Neumann-type computers. Parallel or concurrent architecture will allow machines to do a multitude of operations at the same time. Advanced software designs, VLSI technology, and artificial intelligence technologies will give fifth-generation computers expansive capabilities.

The Japanese, who are devoting massive efforts to AI research, may be the first to exploit fifth-generation technology. Near the beginning of this decade, Japanese industrial leaders decided to launch a national campaign to take the world lead in computer development. At an international conference on computers held in Tokyo in October 1981, Japanese representatives announced their intention to produce fifth-generation computers for commercial use by the 1990s. Edward Feigenbaum, Professor of Computer Science at Stanford University, was one of a handful of Americans invited to the conference. The enormity of the Japanese proposal was immediately obvious to him. If they were successful, the Japanese could replace the Americans as the leaders in computer technology. They could also establish a "knowledge industry" in which knowledge itself would be a salable commodity. "The Japanese," Feigenbaum noted, "understand that if they succeed in this visionary computing project, they will acquire leverage over all kinds of industries, at home and abroad. The Fifth-Generation is an exquisite piece of economic strategy."²¹ Professor Feigenbaum discusses the entire project in The Fifth Generation: Artificial Intelligence and Japan's Computer Challenge to the World, which lie coauthored with Pamela McCorduck. He is concerned that if the United States continues with a business-as-usual attitude, it will squander its technology lead at the rate of one day for each day of delay.

The warnings of Feigenbaum and others have not gone unheeded. The United States responded to the Japanese challenge with a unique new business operation. Austin, Texas, recently beat out more than fifty other cities to become the new home of Microelectronics and Computer Technology Corporation (MCC). This new venture, headed by retired Admiral Bobby Inman (former head of the National Security Agency), is being underwritten by nineteen major U.S. companies. MCC, which represents corporate America's most direct response to the Japanese plan, will concentrate research on software technology, microelectronics packaging, and advanced computer architecture. A Newsweek cover story on the fifth-generation race made the point that the winners will be able to use the new computers to design even more powerful and smarter machines for the future.²² Other nations, realizing the stakes involved, have begun their own fifth generation projects. The Soviets, for example, are joining with their East European allies in a new computer five-year plan to develop expert systems, VLSI microprocessors, improved operating systems, and problem-solving software.²³ Fifth-generation research is critical in every country because of the incredible potential for military and social applications that could ensue.

Artificial intelligence research has progressed significantly in its first three decades. It has grown from a part-time pursuit of a few individuals on the fringes of computer science to a full-fledged field of study. AI researchers now hold international conferences, publish several journals, and collect a sizable share of Defense Department contract money.

From the formative years, through the lean times, and into the present period of popularity, one agency almost single-handedly ensured Al's survival. The Defense Advanced Research Projects Agency (DARPA) supported AI research through two decades of important (and highly risky) research efforts. DARPA's steady support enabled AI researchers to develop the fundamental knowledge and tools that are finally delivering the long-promised intelligent systems. These AI systems already are being used as advisors or consultants in various professional and industrial applications. Artificial intelligence is not a panacea waiting to cure all of our technological problems. It would be foolish, however, for the military not to exploit its full potential. If the promise of this emerging computer technology is fulfilled, commanders will have computer support systems that will enable them to cope more effectively with the formidable challenges that lie ahead.

Center for Aerospace Doctrine, Research and Education
Maxwell AFB, Alabama
and
San Antonio, Texas

1. General Robert T. Marsh, USAF, "A Preview of the Technology Revolution," Air Force, August 1984, p. 45.

2. Avron Barr, Paul Cohen, and Edward Feigenbaum, Handbook of Artificial Intelligence, Vol. I (Stanford, California: Heuris Tech Press, 1981), p. 3.

3. William M. Alpert, "Computers with Smarts," Barron's, 23 January 1984, p. 13.

4. Warren R. Young, "'The Machines Are Taking Over," Life, 3 March 1961, pp. 108-11.

5. Pamela McCorduck, Machines Who Think (San Francisco: W.H. Freeman and Company, 1979), p. 159.

8. Patrick H. Winston, Artificial Intelligence (Reading, Massachusetts: Addison-Wesley, 1977), p. 252.

9. Roger S. Walker, Understanding Computer Science (Dallas: Texas Instruments, 1981), pp. 1-5.

11. Edward A. Feigenbaum and Pamela McCorduck, The Fifth Generation: Artificial Intelligence and Japan's Computer Challenge to the World (Reading, Massachusetts: Addison-Wesley, 1983), p. 17.

12. James A. Saxon and Robert E. Fritz, Beginning Programming with ADA (Englewood Cliffs, New Jersey: Prentice-Hall, 1983). 1), p. 3.

13. Joel N. Shurkin, "Expert Systems: The Practical Face of Artificial Intelligence," Technology Review, November-December 1983, p. 74.

15. William D. Marbach, "The Race to Build a Supercomputer," Newsweek, 4 July 1983, p. 62.

17. J. P. Woodard and E. J. Cupples, "Selected Military Applications of Automatic Speech Recognition Technology," Spectrum, December 1983, p. 38.

18. Raj Reddy and Victor Zue, "Recognizing Continuous Speech Remains an Elusive Goal," Spectrum, November 1983, p. 84.

19. Frederick Hayes-Roth, "Copying Human Knowledge," Spectrum, November 1983, p. 81.

23. Paul Walton and Paul Tate, "Soviets Aim for 5th Generation," Datamation, 1 July 1984, p. 53.

Colonel Pat O. Clifton (B.A., Oklahoma State University; M.A., Creighton University) is Commander, 3480th Technical Training Group, Goodfellow AFB, Texas. He has represented the Electronic Security Command as a Research Fellow at the Center for Aerospace Doctrine, Research, and Education, Maxwell AFB, Alabama; and he has served as Commander, 6993d Electronic Secruity Squadron, Kelly AFB, Texas; Wing Deputy Commander for Operations, Europe; and Chief of the National Security Affairs Department at Air Command and Staff College. His articles have appeared in the Journal of Electronic Defense and the Review. Colonel Clifton is a graduate of both Squadron Officer School and Air War College, as well as a Distinguished Graduate of Air Command and Staff College.

John G. Romo (B.A., Trinity University; M.A., Ph.D., Oklahoma State University) is a mathematician, Directorate of Systems Technology, Hq Electronic Security Command, San Antonio, Texas, and serves as a technical consultant for command applications of artificial intelligence. Dr. Romo was an assistant professor at the University of Texas-San Antonio before serving as a mathmatician in support of the AWACS, Tinker AFB, Oklahoma. His professional experiences and interests include harmonic analysis, Kalman filtering, simulation/modeling, technology infusion, and artificial intelligence.

The conclusions and opinions expressed in this document are those of the author cultivated in the freedom of expression, academic environment of Air University. They do not reflect the official position of the U.S. Government, Department of Defense, the United States Air Force or the Air University.