Expert Systems: A Tutorial

N. Shahla Yaghmai
School of Library and Information Science, University of
Wisconsin-Milwaukee, P.O. Box 413, Milwaukee, WI 53201
Jacqueline A. Maxin
Computer Services, The H. W. Wilson Company, Bronx, NY 10452

Expert systems are intelligent computer applications fashion are often referred to as knowledge-based systems,
that use data, a knowledge base, and a control mech- intelligent computer assistants, or expert advisers [2,4],
anism to solve problems of sufficient difficulty that s i g
nificant human expertise is necessary for their solution.
Expert systems use artificial intelligence problem-solv- The Data Level
ing and knowledgerepresentation techniques to com- On the data level of expert systems is declarative
bine human expert knowledge about a problem area with
human expert methods of conceptualizing and reason- knowledge about the particular problem being solved. It
ing about that problem area. As a result, it is expected is the working memory of the system, the global database
that such systems can reach a level of performance that keeps track of input data for the current problem,
comparable to that of a human expert in a specialized the status of the current problem, and the relevant history
problem area. The high-level knowledge base and asso- of what has been done so far [1,2].
ciated control mechanism of expert systems are in es-
sence a model of the expertise of the best practitioners
of the problem area in question and, hence, human users The Knowledge Base Level
are provided with expert opinions about problems in that
area. Expert systems do not pretend to give final or ulti- On the knowledge base level is procedural knowledge
mate conclusions to displace human decision making; specific to the problem area that the system is set up to
they are intended for consulting purposes only. solve; it is used by the system to reason about the current
problem. Extendibility, simplicity, and explicitness are
basic requirements to consider when choosing a knowl-
edge-representation scheme. The most popular approach
Overview of Expert Systems
is that of forming production rules which become con-
What distinguishes an expert system from an ordinary nected to one other by if-then statements or association
computcr application is that in a conventional computer linkages that build rule networks. Once assembled, these
program, pertinent knowledge and the methods for utiliz- networks can represent a substantial body of knowledge
ing it are all intermixed. In an expert system, the prob- [1,2,5,61.
lem-solving model appears explicitly as a knowledge base Rules in the knowledge base represent both facts and
rather than implicitly as part of the coding, and the heuristics-facts being information that is widely shared,
knowledge base is manipulated by a separate, clearly publicly available, and generally agreed upon; heuristics
identifiable control strategy. Therefore, unlike conven- being intuitive, little discussed rules of thumb that char-
tional computer applications, expert systems incorporate acterize expert-level decision making in the problem
data into the instructions in such a way that new knowl- area. The decision-making power or performance level of
edge can be added to the program without extensive re- an expert system is primarily a function of the sue and
programming [ 1-31. quality of its knowlege base, not of the sophistication of
In short, ordinary computer programs organize knowl- the techniques that it uses to exploit knowledge in order
edge on two levels: data and program. Most expert sys- to draw conclusions and to recommend actions [ 1,7,9].
tems organize knowledge on three levels: data, knowledge
base, and control. Computer systems organized in this The Control Level
The control structure or inference engine of expert sys-
tems makes decisions about how to use the system’s
(019x4 by John Wiley Ji Sons, Inc. knowledge by organizing and controlling the steps it

JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE. 35(5):297-305;1984 CCC 0002-82311841050297-09t04.00
takes to solve current problems. This control mechanism ceeding that of the system; they assist information trans-
is actually a collection of procedures or strategies on how fer in subject areas where human resources are especially
to solve problems [1,2]. scarce and concentrated; and they permit researchers to
discover combinations of facts not previously considered
Total System Operation ~~4,101.
As a result, while the more familiar type of computing
Since most expert systems store problem area knowl- has evolved a single user mode, that of getting answers to
edge as production rules, the knowledge base is referred problems, an ideal expert system has evolved three. They
to as the rule base and the control mechanism called the are: user as client or seeker of answers, user as tutor or
rule interpreter. Rules in the knowledge base are actu- specialist to improve or increase the knowledge base, and
ated by input from the global data base. This rule appli- user as pupil or harvester of the knowledge base for hu-
cation changes the status of the system by modifying the man use. With artificial intelligence programming tech-
global database and by identifying rules pertinent to the niques, an expert system ultimately attains a level of con-
current problem. The rule interpreter finds these perti- sultant expertise higher than that of its tutors, and once
nent rules and decides which to apply via inference proce- specialists collaborate in the development of a system, the
dures or conflict-resolution strategies that match patterns system helps others to develop their own expertise, as well
of rules in the knowledge base with features of the current as to solve problems [ 101.
problems as they exist in the global database. There is no
ordering of rules in the knowledge base; all rules are Existing Systems
evaluated during every cycle as the system tackles current
problem data and keeps track of the status of the develop- Expert systems combine expert human knowledge in
ing configuration [1,2,4]. whatever field is involved with artificial intelligence meth-
ods for the programming of reasoning. Because the con-
trol structure of expert systems is a general reasoning
Hlstory of Expert Systems mechanism, systems can ideally be changed by simply
Artificial intelligence was born as a science in 1956 at adding or subtracting from the knowledge base. Such sys-
Dartmouth College, at a meeting which was instigated to tem architecture provides a convenient way of construct-
discuss ways of simulating thought with computers. This ing sophisticated problem-solving tools that can be used
science now includes not only expert systems, but robo- for many different problem areas. Expert system technol-
tics, speech recognition, image analysis, and the pursuit ogy, first limited to the academic scene, is beginning to
of epistemology, the study of human intelligence in an at- enter into commercial applications. While the predomi-
tempt to formulate rules that explain how humans think nance of systems is in medical and chemical diagnosis,
PI. data analysis, data interpretation, planning, and design,
By the 197Os, it became apparent to the artificial intel- a few pioneering systems exist in a number of other func-
ligence community that inferences or strategies alone, tional areas [1,4,8].
even those augmented with heuristics, were often inade- Following is a partial listing of systems, each followed
quate to solve real lie problems. These problems were by its function and problem area: MYCIN-diagnosis,
usually so complex that, without the addition of more medicine; DENDRAL-data analysis and interpretation,
knowledge about the problem area, either a combinato- chemistry; EL-analysis, electrical circuits; GUIDON-
rial explosion occurred that defied reasonable search computer-assisted instruction, medicine; AM-concept
times, or results were generated that exceeded existing formation, mathematics; VM-monitoring, patient res-
search space limits [l]. piration; RI-design, computer system configurations;
It also became apparent that for many problems, ex- NOAH-planning, robotics; MOLGEN planning, mo-
pert problem area knowledge was more important than lecular genetics; HARPY-signal interpretation, speech
the inferences or strategies used to manipulate that understanding; SACON-user advisor, structural analy-
knowledge. This realization gave birth to the field of sis computer program; XCEL-consultant/intelligent as-
knowledge engineering, which focuses on how to bring sistant, computer sales; IMS-management, automated
expert knowledge to bear in problem solving. This has led factory; +NIX-automatic programming, modeling of
to systems that are able to reason about inferences as well oil well logs: and VISIONS-image understanding,
as about answers to current problems [l]. physics [ 11.
Human needs have weighed heavily in the develop- In addition, a substantial effort is underway to develop
ment of expert systems, particularly the need to save user expert systems for use in the construction of other expert
time and effort in obtaining information. Therefore, un- systems. These include: programming tools (ROSIE,
like more traditional information retrieval systems, ex- AGE, HEARSAY 111, EMYCIN, OPS 5 , RAINBOW,
pert systems provide direct answers to questions, not KMS,EXPERT, ARBY, MECS-A1and UNITS), knowl-
merely citations to documents where answers may be edge acquisition tools (TEIRESIAS, EXPERT, AND
found. In addition, they address the need to train and KAS), and tools for learning from experience (META-
teach new human experts at levels comparable to or ex- DENTRAL and EURISKO). Such systems typically pro-

298 JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE-September 1984

vide formats or languages that represent user interfaces, gence systems. Furthermore, with artificial intelligence
both declarative and procedural knowledge, and control becoming increasingly knowledge-oriented, a substantial
structures for manipulating that knowledge [ 1,2]. portion of these systems could be viewed as having expert
systems components. Nevertheless, the government ap-
Research Sites pears to spend about 10 million dollars a year on expert
systems research and development, and government con-
Expert system development in the United States is tractors using independent research and development
currently being conducted in universities, in not-for-pro- funds probably spend another one to two million dollars a
fit organizations, by the government, and in industrial year in this area [ 11.
settings. While Stanford is the major center of develop-
ment, other universities engaged in research on expert
systems, in descending order of effort, include: the Mas- Construction of Expert Systems
sachusetts Institute of Technology, Carnegie-Mellon For an expert system to be a success, its problem area
University, and about two dozen other campuses. Not- must be well-bounded and nontrivial, with promising
for-profit organizations, in descending order of effort, in- avenues for growth. Ideally, at least one accessible hu-
clude: the Stanford Research Laboratory, the RAND man expert must be found and an appropriate system-
Corporation, the Jet Propulsion Laboratory, and the building tool must be identified. Having met these pre-
Massachusetts Institute of Technology Research Estab- requisites, realistic and incremental objectives should be
lishment. set. Major pitfalls in the development of expert systems are
Government sites include the Naval Research Artifi- poor choice of problems, inadequate resources, and exces-
cial Intelligence Laboratory in Washington, DC first and sive aspirations. Stages of expert system development are:
the Naval Ordnance Systems Command in San Diego, system design, system development, formal evaluation of
CA second. Industrial settings, listed alphabetically in- performance, formal evaluation of acceptance, extended
clude: Advanced Information and Decision Systems, use in a prototype environment, development of mainte-
Amoco, Bolt Beranek and Newman Inc., Bell Labora- nance plans, and system release. So far, no current system
tories, Computer Thought, the Digital Equipment Cor- has completed all these stages [1,7.12.13].
poration, Fairchild, the General Motors Corporation, Although the construction of expert systems has usu-
Hewletl Packard, Hughes, International Business Ma- ally been on the order of 20 to 50 person-years, recent ver-
chines, Intelli-Genetics, JAYCOR, the Machine Intelli- sions of simple systems have reportedly been built in as
gence Corporation, Mart in Marietta, Schlumberger , little as three person-months: however, complex systems
Smart Systems Technology, Systems Control, Teknowl- still require about 10 person-years to complete. Using
edge, Texas Instruments, the TWA Company, and Xerox present techniques, the time for development appears to
Palo Alto Research Center [ 11. be converging on five person-years per system. Most sys-
Japanese companies, such as Fujitsu, Hitachi, and tems require two to five people to construct, and it takes
Nippon Electric are also enthusiastically embracing ex- one to two years for an engineer or computer scientist to
pert system research. In fact, while most American sys- develop into a knowledge engineer [ 11.
tems have around 2000 rules, the Japanese are proposing
the construction of systems on the order of 20,000 rules by The Human Element
the year 1990 via their Institute for New Generation Com-
puter Technology project. Crisis management systems The knowledge base of most existing rule-based sys-
are a high priority item in Japan, and one firm estimates a tems contains hundreds of rules, usually obtained by in-
worldwidc market of 10,OOO systems just for handling terviewing a human expert. Therefore, the human proto-
electric power network crises [I]. type of an expert system must possess special knowledge,
judgment, and experience, and must be able to explain
that knowledge and experience and the methods used to
Funding apply them. Rules and inferences can be gleaned from
So far in the United States, the government has been more than one human, or they can be gathered indirectly
the principal source of funds for work on expert systems. from documents. In such cases, the resultant system rep-
The Defense Advanced Projects Agency and the National resents a consensus of expert thoughts on the problem
Institutes of Health have been the primary funders, fol- area [2,5].
lowed by the National Science Foundation, the Office of Usually, the first body of knowledge obtained from the
Naval Research, the National Library of Medicine, the human includes that of terms and basic concepts, as found
Air Force Office of Scientific Research, the United States in textbooks or journals. However, this is not enough to
Geological Survey, and the National Aeronautics and build a high performance system because human experts
Space Administration. It is virtually impossible to obtain are usually unaware of the exact mental processes by
precise funding figures, not only because expert systems which they diagnose problems. Therefore, an iterative
are not carried as a separate funding category, but be- procedure is followed to improve the system. A sample
cause they are often embedded in other artificial intelli- case is run, during which the human is apt to disagree

JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE-September 1984 299

with the system’s reasoning at some point. This forces the pert system architectures. These approaches are different
human to determine what additional knowledge is needed, implementations of two basic methods for overcoming the
which in turn may elicit a heuristic that can be embedded geometric growth in time and space needed to search’
into precise logical sequences that a computer can exe- complex real life problems. They are: find ways to effi-
cute. As more heuristics are added to the knowledge base, ciently search a space, or find ways to transform a large
the system gradually approaches the competence of the search space into smaller spaces that can be searched
human. Although few existing systems can consistently more efficiently [6].
out-perform a human expert, and although even fewer There is little commonality based on functionality in
can use and effectively integrate the knowledge of several existing system architecture. Instead, expert system de-
experts, the promise is there [10,14,15]. signs are derived from a collection of techniques that deal
with such issues as: solution direction, reasoning with un-
Knowledge Acquisition and Learning certainty, searching small spaces, searching large spaces,
handling large search spaces by transformation or by the
In order to use heuristics, programmers had to develop
development of alternate or additional spaces, and deal-
whole new families of computer languages. The most
ing with time [1,16,17].
common one is LISP, for List Processing, which is espe-
Direction. The commonplace but powerful probleni-
cially proficient at manipulating symbols in addition to
solving paradigm of ru le-based systems involves the
numerical data. However, the key bottleneck in expert sys-
chaining of if-then rules to form a line of reasoning. If
tem construction is the need to have a knowledge engineer
chaining is started from a set of conditions or basic ideas,
interact with one or more human experts in the building of
and moves toward some conclusion, the method is called
the knowledge base. Typically, the knowledge engineer
forward chaining. This has been used in systems for data
helps the human to structure the problem area by inter-
analysis, design, diagnosis, and concept formation. If the
preting and integrating human answers to questions, by
conclusion is known but the path to it is not, as in the case
drawing analogies, by posing counter examples, and by
of goals and hypotheses, the method is called backward
raising conceptual difficulties [ 1,3,6,8].
chaining. This has been applied in systems for diagnosis
To facilitate this process, smart editors have been de-
and planning [16-191.
veloped to assist in entering and modifying knowledge
The problem with forward chaining without appropri-
base rules. In addition, special expert systems have been
ate selection heuristics is that one would derive all possi-
devised to bypass the need for a knowledge engineer by in-
ble outcomes, needed or not. Backward chaining works
teracting with or interviewing human experts directly.
from goals to subgoals by using the action side of rules to
These natural language interfaces model existing knowl-
deduce the condition side of rules. The problem here,
edge bases in order to decide whether or not new informa-
again without appropriate heuristics for guidance, is the
tion fits into what is already known and to use this in-
handling of conjunctive subgoals. In general, this re-
formation to make suggestions to the human. In some
quires finding a case where all interacting subgoals are
sophisticated systems, an explanation module is also in-
satisfied, a search that can result in an explosion of possi-
cluded which allows the user to examine and challenge the
bilities. Thus, in addition to appropriate heuristics, suita-
system’s reasoning process. The end result is the formula-
ble inferences and architectures must be found for each
tion of new rules and the building and augmenting of exist-
type of problem area in order to achieve an efficient and
ing rule sets [4,7,10].
effective expert system [1,16,17].
However, when an appropriate human expert is not
Fortunately, the knowledge of a problem area can help
continuously available-or when the human does not
to guide the steps of the problem-solving process and
have all of the expei-tise desired-self-learning and dis-
therefore the nature of inferences. When the knowledge is
covery approaches to acquiring the needed expertise are
abstract, so that the inference proceeds from abstractions
required. Such tools generally model knowledge bases to
to less abstract statements, the inference is called model
generate, test, and refine a set of candidate rules which
driven. When next steps are chosen by moving from more
are, in fact, discovered by induction from examples. The
to less abstract statements, expectations are being gener-
human enters decisions for each combination of attri-
ated and the inference is termed expectation driven.
butes and values. If the decisions are made by a reason-
However, when working upwards from details or specific
able set of rules, whether known by the human explicitly
problem data to higher levels of abstraction, the inference
or not, the system analyzes the examples and produces
is referred to as data driven. If the next step is chosen on
rules that fit. It also automatically deals with redundan-
the basis of new data or the last problem-solving step
cies and provides rules that contain only the dominant at-
taken, the response is to events and the inference is event
tributes after taking into account inconsistencies among
driven.
the examples [1,4,10].
The event driven inference scheme is similar to for-
ward chaining, except that the conditions or ideas evolve
Sys tern Architecture
over time. Therefore, the next step is chosen either on the
The knowledge engineering community has evolved basis of new ideas or in response to a new situation which
several techniques that can be used to devise suitable ex- has resulted from the last step taken. Although applicable

300 JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE-September 1984

 to many planning problems, the event driven approach is backtracking is referred to as truth maintenance; when
particularly appropriate for monitoring, control, and dealing with incorrect lines of reasoning, backtracking
other real-time operations. is referred to as dependency-directed or relevant back-
The basic control strategies in expert system architec- tracking.
ture may be top down or goal driven, bottom up or data Another approach to large space searching involves
driven, or a combination of the two that uses a relaxation- multiple lines of reasoning that can broaden the coverage
like convergence process to join these opposite lines of of an incomplete search. Here, the chance of discarding
reasoning together at some intermediate point to yield a good solutions from weak evidence is decreased by cany-
problem solution. The hybrid strategy is possible only ing a limited number of solutions in parallel until the best
when the search space is large, or when it can be divided solution can be identified [16,17].
hierarchically so that both forward and backward pro- Transforming Large Spaces. Large search spaces can
cessing can be combined. Such combined searches are be made more manageable by breaking the current prob-
applicable to complex problems that incorporate uncer- lem into smaller noninteracting subproblems. This is pos-
tainties, such as in speech understanding [ 16,171. sible when a number of nonrelated tasks are required to
Uncertainty. Often during expert system develop- achieve a goal. Few significant real life problems fall into
ment, there is uncertainty in data or in knowledge, partic- this category, however.
ularly in systems concerned with diagnosis and data anal- When complex problems can be divided, their sub-
ysis. In such cases, one approach attaches numerical problems usually interact such that valid solutions cannot
values to data and knowledge to indicate the degree of be found independently. Therefore, several approaches
certainty associated with them. Numeric procedures have have been devised to deal with these interactions. With
been devised to handle approximations by combining one method, a fixed sequence of subproblems is found so
evidence. This fuzzy set theory, based on possibilities, that no interactions occur. Another method coordinates
has been used to arrive at a certainty value for the final decision making with available information and moves
solution. the focus of problem solving to the subproblems. Hence,
Another approach involves belief revision or truth decisions are postponed until enough information is
maintenance. Often lines of reasoning are formed from available.
partial or erroneous information. When contradictions A third approach represents the interaction between
occur, incorrect beliefs and the conclusions drawn from subproblems as constraints or relationships that must be
them must all be retracted. To do this, a record of system satisfied, which moves information between subprob-
beliefs and their justifications must be maintained. With lems. By introducing constraints, one is able to pursue a
this approach, redundancies in experimental data can be least commitment style of problem solving. Another ap-
exploited to increase system reliability [ 16,171. proach involves guessing or plausible reasoning. While
Small Spaces. Many straightforward problems, like the least commitment approach is halted when insuffi-
those in design, diagnosis, and analysis, can be searched cient information is available, heuristic guessing can
in small spaces either because the problem is small or carry the solution process along. When wrong guesses are
because it can be broken into small independent sub- made, dependency-directed backtracking can be used to
problems. In other cases, only a single line of reasoning efficiently recover from them.
may be required; therefore, backtracking is not used and As a method to divide complex problems, top down re-
small search spaces are sufficient [ 16,171. finement can prevent combinatorial explosions. With this
Large Spaces. When searching small spaces, exhaus- approach, the most important aspect of the current prob-
tive reasoning by elimination can be accomplished: the lem is abstracted and a high level solution is developed.
system can generate possible solutions and a tester can This limits the search to a small portion of the lower
eliminate solutions that fail to meet certain criteria. In search space, so that solutions can readily be found at
large search spaces, however, a powerful hierarchial gen- each level. A similar approach involves hierarchical reso-
erate and test approach can be effective only if means are lution into contributing subspaces. With this method, el-
available for evaluating possible solutions that are only ements of higher level spaces are composed of elements
partially specified. This approach is appropriate for large from lower spaces, as are syllables, words, phrases, and
data interpretation and diagnosis problems where all sentences in speech understanding. Although these heter-
solutions are desired, since it permits early pruning of en- ogeneous subspaces are fundamentally different from the
tire classes of solutions and thereby reduces the amount top level solution space, possible solutions at each level
of search space required. are useful to restrict the range of a search at adjacent lev-
In the small space generate and test approach, when a els, which also restrains combinatorial explosions.
line of reasoning fails and must be retracted, it can be Large search spaces can also be handled by developing
done by backtracking to the most recent choice point. alternative or additional spaces. When searching a single
However, it is often more efficient to trace errors and in- model is difficult, alternative models for all or part of the
consistencies back to the inference steps that created problem may simplify the search. Or, additional layers
them via dependency records, which is similar to truth can be added to a search space to help decide what to do
maintenance. When dealing with erroneous information, next. These can be thought of as strategy and tactical lay-

JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE-September 1984 301

ers in which meta problem solvers choose among several pected to know or understand the entire system. Users
potential methods for deciding what to do next [16,171. want advice about their current problems and take action
Time. Although expert systems usually do not deal based partly on that advice; in many cases, users are held
with time explicitly, the following approaches have been responsible for the actions they recommend. Therefore,
developed to deal with time in terms of intervals. One in- they must understand the rationale for system decisions.
volves situational calculus, which uses the concept of situ- Knowledge acquisition is an issue because it is difficult
ations that change when sufficient actions have taken and time consuming. The tools developed to date can ex-
place or when new data indicates that a shift is appropri- pand and improve knowledge bases only after a vocabu-
ate. Situations determine the context for actions and, lary and knowledge-representation scheme have been
through the use of frames, can indicate what changes and built. They have no initial specific problem knowledge
what remains the same when actions take place. Another and so they can rely only on general knowledge bases, spe-
approach involves the partial ordering of events. When cific examples of other knowledge bases, and on what hu-
interference between subgoals is observed, the system ad- man experts say about the problem area. Therefore, these
justs the ordering to resolve the interference and, hence, tools must contain, or have access to, the structure, as-
produces a final parallel time-ordered plan [ 16~7,201. sumptions, and limitations of the control mechanism that
will use the new knowledge.
Summary of Expert Systems Limitations Maintaining a large knowledge base is as difficult as
constructing it in the first place. With problems having
The state-of-the-art of expert systems is advancing;
no closed solutions, knowledge bases should change as
however, there are existing limitations as well as potential
human experts accumulate more experience and develop
power. Because of the difficulty in building and main-
new techniques. Maintenance may mean seeking prob-
taining a large knowledge base, the typical problem area
lems in the knowledge base, such as gaps and outdated or
of a system is narrow. Part of the problem is the inability
overlapping items that could lead to inconsistent or redun-
to maintain consistency among overlapping items in a
dant conclusions. A maintenance system, then, should
knowledge base. Therefore, although the optimum con-
have both the syntactic and semantic knowledge needed to
dition is for several human experts to contribute to a sys-
identify and modify items that appear responsible for poor
tem, one human expert or knowledge czar must be on
system performance.
hand to maintain control and ensure quality. As a result,
As expert systems enter into the market place, their
only system builders or other knowledge engineers can
developers will be expected to provide objective evidence
currently operate most systems successfully.
of their validity. Credibility can be increased by making
To further complicate matters, neither knowledge
systems intelligible and understandable so that users can
bases nor control mechanisms contain much information
maintain and modify them. Transparency is required and
about their own assumptions and limitations. They offer
attainable via explanation modules of high-level concepts
little guidance about the appropriateness of new prob-
lems or the boundaries of their own expertise and, so, and familiar models which can tell what systems know,
even the best of them come up with wrong answers. Fur- how systems use their knowledge, and why systems reason
thermore, the available knowledge-representation lan- as they do. For systems to have high performance, careful
guages are limited,input-output languages and inter- attention must be paid to the representation of knowl-
faces are relatively inflexible and stylized, and system edge, the methods of inference, and to performance vali-
explanations are equally limited and stylized. dation. For systems to have utility, a large body of knowl-
All of the above contribute to laborious knowledge
edge about a sizable problem area is required, which
base construction. The problem of automating knowl- requires careful attention to knowledge acquisition and
edge base construction has been considered, as well as the knowledge base maintenance.
writing of routines that carry on dialogues with humans in Presently, it is difficult to formulate precise questions
order to elicit knowledge without the aid of a knowledge that can be answered experimentally. However, a taxon-
engineer. So far, however, these activities are successful omy of problems and solutions is being worked out which
only when the system contains an initial framework to will make it possible to develop criteria to determine the
build on [21]. best method of solving a given problem. Since expert sys-
tem construction and experimentation is expensive, there
Issues in Expert Systems is a trend toward design tools that build systems within a
given framework and, hence, speed up the construction,
Increased emphasis on large knowledge bases have modification, and experimentation with existing systems.
made explanation, knowledge acquisition, and mainte- The last outstanding issue is that of choosing the right
nance and validation, critical issues for expert systems. In framework for solving a problem before searching for a
addition, since the importance of expert systems was not solution. It can be viewed as two stages: choosing a lan-
fully anticipated, there is now more interest in experi- guage, and selecting the best solution within it. It is now
mentation and more emphasis on choosing an appropri- possible to understand how to use heuristics to find and
ate framework for problem solving. select solutions to problems within a given problem-solv-
Explanation is needed because users cannot be ex- ing framework; then they will be useful for theory con-

302 JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE-September 1984

struction and hypothesis formation within existing theo- these developments may very well herald a maturing in-
ries [20]. formation society where expert systems put expert knowl-
Other issues include the lack of adequate and appro- edge at everyone's disposal. As production and informa-
priate hardware, and the need to improve and expand the tion costs decrease, major new opportunities should open
capture of approaches and architectures so they can be up for the social betterment of all [l].
drawn together and synthesized into a firm base upon
which to build future expert systems. Continuous funding
and more knowledge engineers are needed, as well as an Implications For Information Science
orderly method for taking successful research projects Inference making, knowledge creation, and machine
and translating them into the huge number of potential translation have been discussed in library literature at in-
expert system applications [ 11. tervals over the last three decades. In 1952, the possibility
In short, research is required to develop: improved of using machines to manipulate information to make
knowledge acquisition systems, better explanation sys- new discoveries was a subject of controversy among li-
tems, learning by example, friendlier user interfaces, brarians. However, by 1965, it was suggested that librari-
more [Link] knowledge engineering tools, better sys- ans could help computer scientists create systems capable
tem architectures and inference procedures, more effi- of inference by end users by refining the techniques of
cient techniques for working with multiple human ex- concept representation. In fact, it has been claimed that
perts, more adequate methods for dealing with time, the automatic theory construction would be the result if the
ability to make appropriate assumptions and expecta- procedures used by humans to construct them could be
tions about the world, the ability to exploit casual models discovered and embodied in standard semantic pro-
and couple them with other knowledge, general methods cedures [22-25).
for planning, analogical reasoning, methods for coupling These techniques of concept designation have also
formal deduction into expert systems, parallel processing been considered for use in translation: first, via ma-
approaches, and better knowledge-representation meth- chines-a hypothetical common language base of human
ods. In addition, substantial research is needed to de- communication-then via the infeasible idea of using the
velop expert systems for problem areas where there are no Colon Classification as an international language. Al-
human experts [21]. though this early interaction between machine transla-
tion and information retrieval was not particularly fruit-
ful, it must be noted that the PRECIS indexing system
Trends in Expert Systems
imposes a structure in entries that make them amenable
Since there appears to be no functional limitation to to machine translation [21,26-281.
the ultimate use of expert systems, they will most likely The direct application of expert systems to informa-
find eventual use in almost every endeavor where sym- tion work has been recent, primarily in the area of online
bolic reasoning with detailed professional knowledge is bibliographic database searching. On the basis of early
required. In the process there will be exposure and refine- trials, there is a strong indication that future users will be
ment of the heuristics of the various fields of expert sys- able to conduct comprehensive and successful online lit-
tem application. This addition to human knowledge will erature searches without the aid of a librarian or informa-
undoubtedly be the most important contribution of the tion specialist [29].
knowledge-based systems approach. Existing systems in the United States include: PA-
In the short term, however, expert systems with thou- PERCHASE, a self-servjce facility that requires no for-
sands of rules will appear. In addition to an increase in mal instruction, user manual, or trained human interme-
rule-based systems, there will be an increase in the num- diary, thereby reducing the cost of searching; FRED,
ber of non-rule based systems, since not all problem areas which fulfills connection, communication, and central-
are homogeneous enough to be readily cast into the pro- ized processing functions; IIDA, which instructs users
duction system framework. Improved explanation sys- and provides them with diagnostics; and OL'SAM which
tems will be capable of telling why a system does what it provides front ends to databases. In addition, the Na-
does and what things are of importance. tional Library of Medicine is developing the Hepatitis
By the late 1980s, intelligent, friendly, and robust hu- Knowledge Base which provides answers to questions in
man interfaces will be available, as well as better system the form of authoritative statements, thereby bypassing
building tools-even for developing systems in areas the lengthy reference retrieval, document acquisition,
where there are no human experts. By 1990, there will be and perusal cycle [9,23,30-321.
knowledge acquisition systems which, after being given Although little progress has been made in developing
the content of a basic problem area, will rapidly guide a expert systems for traditional library work, the rule-
human in forming the needed knowledge base. In the based activities of cataloging and classification seem par-
199Os, expert systems will begin to proliferate in areas ticularly worthy of attention. In the area of cataloging,
where there previously were no human experts. two scenarios suggest themselves: a first-step person-
Then,around the year 2000, systems will semi-autono- machine interface where intellectual effort is divided be-
mously develop knowledge bases from text. The result of tween the human intermediary and the support system;

JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE-September 1984 303

and an expert system with full cataloging capability linked 6. Davis, R.; Lenat, D. B. Knowledge-Based Systems in Artificial
to an electronic publishing system so that as text is gener- Intelligence. New York: McCraw-HUI; 1983.
7. Duda, R. 0. “Knowledge based expert systems come of age.”
ated online, it can be passed through the cataloging pro- Byte. 6(9):238-281; 1981.
cess without intellectual input from humans. A study at 8. Feigenbaum, E. A. Knowledge Etigineeringfor the 1980‘s. Stan-
Exeter University in the United Kingdom will investigate ford, CA: Stanford University, Computer Science Department;
the use of AACRZ and the MARC Manual as sources of 1982.
rules for a cataloging knowledge base as well as natural 9. Mareus, R. S. “An Automated Expert Assistant for Information
Retrieval.” In: L. F. Lunin, M. Henderson, and H. Wooster,
language data querying, intellectual storage, and retrieval Eds. The Infurmation Community: An Alliunce fbr hugress:
from the database 1331. Proceedings of 44th ASIS Annual Meeting. White Plains, NY:
An expert system classification system can also be en- Knowledge Industry Publications; 1981: 270-273.
visioned, with a natural language exchange which could 10. Michie, D. Knowledge-Bused Systems. (Report 80-1Mll) Ur-
sift, weigh, and rank order terms of greatest use and rele- bana-Champaign: University of Illinois; 1980.
11. Feigenbaum, E. A.; McCorduck, P. The Fifth Generation.
vance before conducting a search through its knowledge Reading, MA: Addison-Wesley; 1983.
base to match retained terms with subdivisions. This ex- 12. Davis, R. “Expert systems: Where are we? and Where do we go
ercise would be followed by a rationalization exercise be- from here?” A1 Magazine. 3(2):3-22; 1982.
fore the optimal notation is decided upon 1241. 13. Hayes-Roth, F. Al: The New Wuve-A Technical Tutorial Jbr
Doubtless, other rule-based tasks could be enhanced R%D Management. (AIAA-81-0827) Santa Monica, CA: Rand
Corporation; 1981.
by the expert system approach including that of indexing 14. Lenat, D. B.; Sutherland, M. R.; Gibbons, J. “Heuristic search
and abstracting. Knowledge bases could also be con- for new microcircuit structures: An Application of artificial intel-
structed to consult in the areas of hardware and software ligence.“ Al Magazine. 3(3):17-33; 1982.
selection and evaluation, and in the areas of information 15. Marcus, S. J. “Computer systems applying expertise.” Zhe New
system design and evaluation. Yurk Times. 132 (45785): Dl-D3; 1983.
16. Heuristics Programming Project 1980. Stanford, CA: Stanford
Such developments not only have implications for edu-
University, Computer Science Department: 1980.
cation for library and information science, but also for 17. Stefik, M.; Aikins, J.; Balzer, R.; Benoit, 1.; Birnbauni, L.;
the role of the information professional. The use of expert Hayes-Roth, F.; Sacerdoti, E. “The organization of expert sys-
systems may well take away from current professional tems, A tutorial.” Artificial Intelligence. 18(2):135-173; 1982.
functions but it may in turn create new roles, or at least 18. Alexander, T. “Practical uses for a useless science.” Furtune.
1W11):138-145; 1982.
release the professional to carry out other tasks and func-
19. Bramer. M. A. “ A Survey and Critical Review of Expert Systems
tions. The major focus of controversy has been the dis- Research.” In: D. Michie, Ed. Zntrudiictury h’cwdings in Expert
tinction between providing information and giving advice Systems. New York: Gordon & Breach; 1982: 3-32.
which requires more knowledge, hence opening the door 20. McCarthy, J.; Hayes, P. J. “Some Philosophical Problems from
to personal bias. However, at this point we must under- the Standpoint of Art al Intelligence.” In: B. Metzler and D.
Michie, Eds. Muchine Intelligence 4 . New York: Halsted; 1969:
score the more important role of the information profes-
463502.
sional-that of a gatekeeper in the area of information 21. Buchanan, B. G. “New Research on Expert Systems.” In: J. E.
service. To accept that role is to know that the assimilation Hayes and D. Michie, Eds. Machine Intelligence 10. New York:
or rejection of new concepts and technologies will not only Academic; 1982: 269-299.
shape the future of information services, but determine 22. Farradane, J. “The Necessity For Semantic Analysis of Informa-
tion Retrieval.” In: Infurmutics l: l’ruceedings ?/’ u Cor!feretrce
the progress of our civilization and culture 1271.
Held at Durham University 1973. London: Aslib; 1974: 67-74.
23. Meadow, C. -1.; Hewlett. T. B.; Auersa, E. S . “A Computcr ln-
termediary for Interactive Data Base Searching/Design.” Jorirnal
Acknowledgement uf’ the Americtrn Suciet,~f o r Infurniutiun Science. 33(5):325-
332; 1982.
The authors wish to thank Mr. Gordon S. Black, Jr.,
24. Wright, W. E. “The Subject Approach to Knowledge: Historical
Senior Research Associate at the PPG Fiber Glass Re- Aspects and Purposes.” In: M. F. Tauber, Ed. The Siibject
search Center in Pittsburgh, PA for the time that he Analysis of Library Materials. New York: Columbia Univcrsity
spent critiquing the manuscript. School of Library Service; 1953: 8-15.
25. Sharp, J. R. Some Frtndumentuls of’Inf’ornirrtiun Retrieval. New
York: London House & Maxwell; 1965.
References 26. Austin, D.; Sorenson, J.; Lanibert, C. “PRECIS in a multilingual
context.” Lihri. 26( I ): 1-32, 26(2): 108-139, 26(3): I81 -2 15. 26(4):
1. Gcvartcr, W. B. An Owriiew o/’ Expert Systems. (NBSIR 82- 302-324: 1976.
23)s) Washington. DC: National Bureau of Standards: 1982. 27. Weaver, W. “Translation.” A 1940 Memorandum In: W. W .
2. Nail. D. S. “Expcrt computer systems.” Computer. 16(2):63-85; Lmke and A. D. Booth, Eds. Muchine Trartsblir~n~JJ’Ltm-
1983. p a g e s . Cambridge, MA: MIT Press; 1955: IS-23.
3. Shurkin. J . N. “Expert systems: The practical face of artificial in- 28. Ranganathan, S. R. Nalegonrenu to Librury ChSif~[Link]-
tell igcn cc . T e c h n o l u ~Review.
” ~ 86(6): 72-78; 1983. don: Library Association; 1957.
4. Hertz. D. “Artificial intelligence and the business manager.” 29. Clark, A.; Cronin, B. “Expert systems and library/inforniation
C w t r p t e r World. 17(43):19-26; 1983. work.” Journal of Librariatrship. 15(4):277-292; 1983.
5. Buchanan. B. G . ; Duda. R. 0. “Principles of Rule-Based Expert 30. Bernstein, L. M.;Siege], E. R.; Goldstein, C. M. “The hepatitis
Systems.” In: M. Yovits, Ed. Advutrces in Cunrpriters. New knowlcdge base: A prototype information transfer system.” An-
York: Academic; 1983: 164-21 7. nals of Internal Medicine. 43( 1): 165-222: 1980.

304 JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE-September 1984

31. Horowitz, G . L.; Bleich, H. L. “Paperchase: A computer pro- graphic information retrieval.” Information Technology and Li-
gram to search the medical literature.” New Englund Journal of braries. 1(4):317-326; 1982.
Medicine. 305(16):924-30; 1981. 33. Davies, R. “Document, information or knowledge?” Journal of
32. Taliver, D. E. “OL‘SAM: An intelligent front-end for biblio- Librarianship. 15(1):47-65; 1983.

JOURNAL OF THE AMERICAN SOCIETY FOR INFORMATION SCIENCE-September 1984 305