Guest Editors' Backward and forward chaining-our nascent science of

Introduction Artificial Intelligence-has ancient antecedents: Aristotelian

rhetoric reasons deductively from the Universal to the Particular;
Socratic dialectic reasons inductively from the Particular to the
Universal, as does scientific method. The seedling confluence
of computer graphics and Artificial Intelligence has deep roots.

Computer Graphics
5nd Expert Systems
Norman 1. Badler and Timothy W. Finin
University of Pennsylvania

As computer graphics evolves and expands into an problem: How does the three-dimensional world appear
essential component of many interactive systems, it is when projected onto two dimensions. Another relation,
almost inevitable that the applications themselves become one shared by robotics, is that graphics provides an
more complex and ambitious. In particular, the interactive important tool for presenting the data being processed by
system is rapidly becoming the intelligent system: Not only these systems as well as their internal states.
does the interface permit convenient access to the entities Other areas within Al also have a natural relationship
and operations of the application, but it also aids and with computer graphics. Natural language processing, for
augments the mere interaction in significant ways. For example, was once seen as offering a kind of interface that
example, computer-aided design systems check topological would be an alternative to a graphic one. The current
constraints while an object is being constructed, robot trend, however, is to build sophisticated interfaces that
simulators assist in avoiding collisions, and slide-making combine both graphic and linguistic modalities. The rela-
systems have automatic layout modes. tionship between graphics and natural language processing
The implication of these aids is that the new generation becomes more interesting when one looks at the underlying
of graphical systems will provide increasing intelligence in issue-communication. The study of natural language
the human-computer interface. The study of the charac- processing is really the study of communication between
teristics and capabilities necessary to effect communication intelligent agents. In dealing with this larger problem of
between human and machine on this higher level is a communication, researchers have developed a number of
central aspect of Artificial Intelligence (AI). One of our theories and techniques directly relevant to systems that
principal goals in presenting this special issue is to bring the communicate with people through graphics as well.
approaching confluence of computer graphics and artificial The papers presented in this special issue of IEEE
intelligence into sharper focus. In addition, we hope to spur CG&A involve the interaction of computer graphics and
interest in common issues and open new pathways for the relatively newer subfield of Al, expert systems. Each of
future graphical systems. the three systems described in this special issue is offered as
The fields of artificial intelligence and computer graphics a graphic interface to an expert system. Moreover, each
have interacted in a number of ways in the past. This system, although primarily designed as a graphic interface,
interaction has been particularly strong in the areas of makes good use of some of the new techniques that have
image processing and robotics. Image processing and come out of Al research. Thus, each can be characterized
graphics are naturally related because they share a common as an expert system.

November 1985 A.7

v2/2-1 XI1- .-
I., 1 IIH)/O.II-
. .. .. .. .V
02 SW/ VV-' 025$01
_V.W ei
wn
W_ 19S5
i oQ,;
-
s70 IEEE
1 25
Expert systems 5. Explanations. When appropriate, an expert system
should be able to justify its conclusions. This can be
An expert system could be defined as a computer done by offering an explanation which describes the
program which solves problems in some particular domain; system's reasoning that led to the conclusions. It may
these problems normally require some specialized knowl- also be appropriate for some systems to explain why
edge (e.g.,expertise) on the part of humans who successfully other plausible conclusions were not reached.
solve them. The operant terms in this definition are solves
problems, particular domain, and specialized knowledge. 6. Shell architecture. An expert system's typical archi-
This definition would exclude programs that are not
tecture is a )shell, which includes a number of general,
characteristically problem solvers (e.g., a text editor), problem-independent components, combined with
attempt to be very general (e.g., a theorem prover), solve
one or more specific knowledge bases encoding the
problems that are considered straightforward implemen- problem-specific information. The general compo-
tations of well-defined theories (e.g., computing fast Fourier nents can include an interaction manager, a general
transforms), or are just mundane (e.g., simple record- "inference engine," a knowledge-acquisition system, a
keeping). Typical domains for expert systems are medical KB debugger, a KB editor, and an explanation
diagnosis, signal interpretation, fault diagnosis, and com- generator.
puter configuration.
Unfortunately this definition is still much too broad: It None of these features is required to characterize a
encompasses most of what used to be called "application
system as expert. For example, many Al problem-solving
programs." To capture what is new and valuable in recent systems do not use a rule-based approach, and explana-
work on expert systems, it is useful to enumerate some of tions are not relevant in some problems. All these features
the major features found in "good" examples of expert have been found quite useful in building powerful prob-
systems. In addition to the partial definition offered above, lem-solving systems, however.
a good example of an expert system would be likely to
have the following characteristics:
The origin of expert systems
1. Separate knowledge base. The special domain
knowledge that the expert system uses is explicitly Our current notion of an expert system as a natural class
represented in a module (the knowledge base or KB) of computer programs was first articulated in the early
which is separate from the components that use it. 1970's by Edward Feigenbaum and his colleagues at the
2. Multiple use of knowledge. Since the KB is explicitly Heuristic Programming Project (HPP) of Stanford Uni-
represented as a separate module, it can be used in versity. Their work on problem solving in Al was a
several different ways: to make decisions, to construct methodological departure from preceding research, which
explanations, to construct tutorials, etc. This requires had focused on the development of general problem-
that the knowledge be represented in a more general solving strategies that could be used in conjunction with
way so that it does not favor one use at the expense of
detailed descriptions of a particular problem to produce a
another. The goal is to express the knowledge in a solution. Feigenbaum and his colleagues at HPP accepted
general way that allows it to be reasoned about as the legitimacy of representing and using as much problem-
well as reasoned with. specific knowledge as possible. Thus they de-emphasized
3. Special knowledge representation languages. The the attempt to discover and use general principals of
knowledge in an expert system is usually encoded in a intelligence and accepted the utility of building in problem-
special representation language. Most current Al specific expertise.
systems use representation languages that rely on one The first system this group produced was called Heuristic
or more of three general techniques: rules, frames, DendraL This system attempted to identify likely chemical
and logical relations. structures for unknown compounds based on their chemi-
4. In the knowledge lies the power. This often-quoted cal formulae and data from a mass spectrogram. Besides
slogan underscores the fact that the real intelligence acknowledging the need for domain-specific expertise,
in the expert system lies in the domain knowledge Dendral pioneered the representation of this knowledge in
represented in its KB, not in any of the more general the form of rules-a technique that has turned out to be
components. As a corollary, any search strategies for important for many subsequent expert systems.
problem solutions should be expressed in the general This technique was further developed by the HPP group
KB rather than in the code that uses the KB. in the expert system MYCIN, which assisted physicians in
26 IEEE CG&A
choosing appropriate therapies for patients suspected of The articles
having bacterial infections. Besides encoding its knowledge
in the form of rules, the MYCIN development made In the first article of this special issue, Steve Feiner
important advances in the areas of reasoning with uncertain describes a system (APEX) that generates pictures portray-
facts and rules, generating explanations for conclusions, ing the performance by a problem solver of physical
managing a consultation dialogue, and providing environ- actions in a three-dimensional world. As a graphic system,
ments for knowledge acquisition. APEX relates to the notion of an Al expert system in two
ways: First, APEX was designed to generate pictures from
the output of an expert system that plans physical actions.
Second, APEX is a kind of expert system itself. It takes a
Rule-based systems general description of a set of physical objects, an action to
be performed on them, and a simple model of a human's
In simple terms, a rule-based system has two major knowledge of the scene and determines what would be a
components: a knowledge base and an inference engine. good graphic description of the planned action for that
The knowledge base contains a set of facts and a set of person. To do this, it must decide what objects should be
rules, which together represent the system's general domain included in the picture, how much detail must be used, how
knowledge and specific knowledge about the current the action can be indicated, and a host of other factors.
problem. A fact is generally considered to be an atomic The second article describes a system called SAGE,
proposition which is true in the world, and a rule is developed by Eric Clemons and Arnold Greenfield. Their
considered a conditional with an antecedent (if) part and a system is designed as a general framework for building
consequent (then) part. Thus a rule is knowledge of the graphic interfaces for a wide class of sensitivity analysis
form: models-models that certainly fall within the general
definition of an expert system. The authors make the point
IF/ antecedent-THEN/ consequent that sensitivity analysis systems require their users to make
changes in the model and quickly evaluate the results. The
Depending on the nature of the antecedent and conse- graphical presentation of the effects of a model change aid
quent parts, a rule can be one of several different varieties. the quick evaluation.
If both parts are viewed as logical propositions, then the The final article describes GUIDON-WATCH, a graphic
rule can be viewed as a rule of inference. If the antecedent interface for browsing and viewing a class of expert
is interpreted as a partial description of some state of the systems based on NEOMYCIN. The authors have included
factual knowledge base and the consequent can be any a number of techniques for displaying the various compo-
executable expression, then the rule is similar to the notion nents of a large, complex expert system. GUIDON-
of a production rule as used by cognitive scientists. Many WATCH exploits trees, tables, icons, animation, and other
rule-based systems have a need to deal with uncertain data graphic formats in a window-based environment. The
and rules which do not always hold. A common variation, result is an interface that helps both the end-user and the
then, is to associate with each fact or rule a degree of system designer understand, debug, and modify a large
certainty which represents the system's confidence that the expert system.
fact or rule is true.
The inference engine is a kind of rule interpreter which
takes the set of facts and rules and computes additional
facts or rules that hold. An inference engine can use rules in
one of two important ways. Forward chaining reasons For more information
from the initial set of facts to derive additional facts which
must hold. The inference engine identifies rules whose Some sources for more information on artificial intelli-
antecedent parts are satisfied by the facts currently in the gence, expert systems, and knowledge representation
KB. Backward chaining reasons from a given goal back- follow:
ward to a set of facts which, if in the current KB, would
support the goal. Given a goal to satisfy, the inference "Special Section on Architectures for Knowledge-Based
engine looks for an appropriate fact or rule in the KB. If a Systems," Communications of the ACM, Vol. 28, No. 9,
matching fact is found, then the goal has been satisfied. If a Sept. 1985.
rule is found in which the consequent matches the goal, This issue contains three tutorial articles on the three
then the engine attempts to satisfy the rule's antecedent dominant paradigms for building expert systems-frames,
recursively by setting it up as a subgoal. rules, and logic.
November 1985 27
Brachman, R., and H. Levesque, eds., Readings In Knowl- Kowalski, R., Logic for Problem Solving, Elsevier, New
edge Representation, Morgan Kaufman Publishers, Los York, 1979.
Altos, Calif., 1985. This classic book presents the use of logic as a language
This book contains about 30 seminal articles on various for representing knowledge and solving problems.
issues concerning the representation of knowledge. These
articles are representative of most of the important ideas
developed cver the last 15 years. Shortliffe, E., and B. Buchanan, Rule-Based Expert Sys-
tems, Addison-Wesley, Reading, Mass., 1984.
Charniak, E., and D. McDermott, Introduction to Artifi- This edited collection of articles, based on research done
cial Intelligence, Addison-Wesley, Reading, Mass., 1985. at the Heuristic Programming Project at Stanford, dis-
This is a good general introduction to Al, emphasizing cusses the important expert system MYCIN and its de-
the use of logic and deduction. scendants, as well as general issues.

Hayes-Roth, F., D. Waterman, and D. Lenat, eds., Build-

ing Expert Systems, Addison-Wesley, Reading, Mass., Waterman, D., A Guide to Expert Systems, Addison-
1983. Wesley, Reading, Mass., 1985.
This book contains articles that address both theoretical This book attempts to introduce and explain expert
and practical issues. An interesting feature of the book is systems to readers who do not have a background in Al. It
that it presents a typical expert system problem and shows contains a comprehensive catalog with descriptions of over
how a number of different expert system building tools 200 expert systems.
could be employed to solve it.
Weiss, S., and C. Kulikowski, A Practical Guide to
Special Issue on Knowledge Representation, Computer, Designing Expert Systems, Rowman and Allanheld,
Vol. 16, No. 10, Oct. 1983. Totowa, NJ, 1984.
This special issue of Computer gives a good picture of This book is based on the experience of Weiss, Kuli-
the current state of the art and current research areas in kowski, and their colleagues at Rutgers University, where
knowledge representation. they built expert systems using the system EXPERT.

Timothy W. Finin is an Assistant Professor of

computer and information science at the
Norman I. Badler is Associate Professor of University of Pennsylvania, where he has
computer and information science at the Uni- been a faculty member since 1980. His re-
versity of Pennsylvania, and has been on that search interests include computational lin-
faculty since 1974. Badler is also a coprincipal guistics, knowledge representation, and expert
systems. In 1983 he was awarded an IBM
investigator of research projects for the faculty development award.
National Science Foundation and has a US Finin received the BS in electrical engi-
Army Research Office artificial intelligence
grant. neering from the Massachusetts Institute of Technology, and the
Badler was named associate editor-in-chief MS and PhD in computer science from the University of Illinois.
of IEEE CG&A earlier this year, and has been an active He did computer vision research at MIT's Artificial Intelligence
participant in ACM Siggraph since 1975. He was conference Laboratory from 1971 to 1974, and has worked in natural
tutorial chair from 1976 to 1979, and was elected vice-chair in language processing with the Coordinated Science
1979 and again in 1981. He is the coauthor of more than 35 Laboratory.
technical papers, and operates a computer graphics research
facility with full-time staff and about 25 student participants.
Badler received his BA in creative studies mathematics from
the University of California-Santa Barbara, and his MS in The authors' address is University of Pennsylvania, Computer
mathematics and PhD in computer science from the University of and Information Science, School of Engineering and Applied
Toronto. Science, Moore School D2, Philadelphia, PA 19104.

28 IEEE CG&A