-5.1- WHAT IS ARTIFICIAL INTELLIGENCE
The term artificial intelligence (AI), coined by John McCarthy in 1956, is a single term used to describe a collection of programming techniques which are intended to solve problems that cannot be solved by conducting an exhaustive search of the alternatives (the brute force technique) because of the large number of alternatives (the combinatorial explosion). This approach is not suited to the sequential and numerical computers (the von Neumann machines) which are in common use and so the techniques have remained academic curiosities. The development of microcircuitry has brought the non von Neumann, or parallel, machine into the realm of possibility. The computer which can process logical data in parallel is sometimes referred to as a "fifth generation" computer.
In Its short history the computer has been used for the routine tasks which human beings perform slowly and with frequent errors. The instructions for the computers, like the instructions for the humans they replaced, were detailed and difficult to change. The AI techniques provide flexible instructions for the computers, allowing the computers t6 adapt to change without human intervention. The goal in the design of the new machines is to replicate the higher level human thought process to solve problems that cannot be solved with the "brute force" approach. This naturally leads to the problem of defining what constitutes the human thought process.
AI research has been divided into two groups. One group believes that AI should imitate the human cognitive process which occurs when one is thinking to oneself. This process is a network of rules of thumb (heuristic rules) based on the imperfect information obtained through experience. The other group believes that AI should imitate the reasoning processes that humans use when communicating with each other. The fragmentation of the AI programming techniques has occurred over exactly which process constituted the thought process and should become the main goal of AI research. Human cognition has been the subject of psychological research for many years and there is still no consensus among the experts in the field. It is not surprising then that the two areas of research in AI have coexisted for over thirty years and there's no resolution in sight. The result for AI is that no unified theory has evolved.
There is additional confusion when AI researchers discuss research problems or evaluate each other's research and neglect to spell out which theoretical approach is being used in a particular project. A program successful in the eyes of one AI researcher whose only intention is to catalog and utilize the informal rules of a human decision maker will not be considered successful by another AI researcher whose criterion for success is arriving at a conclusion using a reasoning process.
Paul Roller [7 ch. 4] compares the field of human factors engineering for computers to Frederic Taylor's work with the physical capabilities of humans. He expresses the decisions to be made as essentially the redistribution of labor. The AI technology brings into question exactly which activities are best assigned to human beings. "The human-computer interface will become more critical and sophisticated" [7 ch.4 p. 90] after the labor of understanding is shifted from the user to the computer. A review of the range of AI programs in current use to solve non numerical problems is provided by Barr [5 and 6]. The two most promising AI applications currently in use have been natural language (NL) systems and expert systems. With the use of these systems, the burden of understanding is being shifted from the users to the computer. The degree to which this shift is achieved is described further in the following sections.
-5.2- WHAT IS A NATURAL LANGUAGE SYSTEM
Natural language is the spoken medium in which people communicate with each other. Natural language contains ambiguities that help human beings communicate complex concepts which would require too much time and effort to communicate in detail: e.g., large, tall, cold, etc.. These concepts take their meaning from the context in which they are used. Human beings are able to understand the ambiguities by applying some common sense reasoning. Human beings are also very flexible and can carry on several such tasks simultaneously (e.g., "composing sentences and recalling editor commands, while at the same time, the person may be thinking about a program compilation that was just completed, and be interrupted by a phone call" [7 pg. 18]). These human capacities are not easily delegated to the computer, particularly if the computer must react rapidly on the basis of incomplete information. However, within a narrow range, computers can take on the task of understanding ambiguous language.
Computers have traditionally been structured to carry out only carefully planned and unambiguous instructions (the programming language). These instructions in the usual von Neumann architecture have to be completed in the strict order (or procedure) specified by the programming language. This has allowed the computer to automate many routine tasks previously performed by humans, but at the same time it has limited the uses to which computers can be put. Computer architecture has become more flexible with the development of the VLSI (very large integrated circuits) that allow the development of a computer which can process instructions simultaneously (in parallel). This in turn makes it possible for computers to work directly with untrained (non programmer) users. New programs are being developed to allow computers to deal with the ambiguities of natural language: these programs are referred to as NL systems.
Many NL systems are currently available [37] and their domain of expertise is usually related to a database management system (DBMS). The ordinary DBMS requires a user to be able to use a query language and to have some knowledge of how data are stored. Moreover, a detailed knowledge of the structure of the database may be needed to design a query which makes efficient use of the DBMS. This requires a technical expertise which is not a prerequisite for being able to efficiently use the information after it is retrieved. The natural language systems currently available allow a user to use English phrases to query the system. The English sentences are translated into the query language appropriate to the DBMS and structured to make efficient use of the system. Another application for NL systems is as an aid for translators who "supervise" the translation of documents from one language to another by the NL system. These human translators use the NL translator for the majority of the translation and make translations of only those passages which the NL system is unable to handle.
Both the DBMS and language translation systems illustrate the function of the NL system. The NL system allows non programmers to effectively use the capabilities of computers without having a detailed knowledge of how the computer is working (this is referred to as the transparent system). The object of the NL system is to "define a human-engineered subset of natural language that would place a much smaller burden on the computer.... Specifically, this language subset should be easily and quickly learned even by computer-naive users, easy to use and highly versatile once it is learned, neither intimidating nor frustrating for the user, and not easily forgotten even after long periods of time." [7 p. 81] Most NL systems allow the user to define his own dictionary of words which are meaningful to both he and the computer.
The NL system that can handle every English word and phrase does not yet exist. Neither the completely parallel machine nor the programs necessary to handle all the ambiguities of NL exist yet. Still, great progress is being made in these areas and the intermediate NL systems are being applied profitably in business applications.
-5.3- WHAT IS AN EXPERT SYSTEM
The human expert possesses intimate knowledge of a small area which allows him to make judgments based on his specialized knowledge. Human experts are frequently promoted, transferred or they retire and their specialized knowledge of a particular area is lost and must be redeveloped by a replacement. Human acquisition and updating of knowledge is inefficient because the effort necessary to develop the expertise must be repeated over and over again.
The computer expert system either mimics or supports the decision maker. The area of expertise that can be represented in an expert system is very narrow, but has the advantage over human experts in that it can be copied (cloned) as often as needed as long as the necessary hardware is available. The flexibility and narrow domain of knowledge have recommended commercial expert systems to diverse applications. The current list of commercial applications revolves around two areas: the diagnosis of a problem and the specification of some action to be taken to deal with the problem. The problem orientation of the systems makes them ideal for keeping the expertise necessary for special problems close to where the problem is located.
Another limitation of the computer based expert system is that it is limited by the knowledge of the expert on which the system is based. The expert system is created by questioning a human expert about the area of his expertise, asking him about rules of thumb (heuristic rules) that he uses to make decisions. These rules become the knowledge base of the expert system. The importance and certainty that the human expert places on the rules determine how the rules are applied in the expert system. Since this type of expert system cannot generate its own rules, changes in the rules of thumb the expert uses must be entered into the expert system to keep it current.
Clearly the expert system does not replace the human being. It can make judgments similar to the judgments of human experts: however, it cannot make the common sense judgments that the human expert uses to fabricate the rules of thumb (heuristic rules) for making the judgments that result in the changes in the rules he uses. There is as yet no theoretical bridge between the techniques used in an expert system and common sense reasoning, although some AI researchers make this a long term goal.