14 IEEE JOURNAL OF OCEANIC ENGINEERING, VOL. 13, NO.

2, APRIL 1988

Artificial Intelligence-Definition and Practice

ASA B. SIMMONS AND STEVEN G. CHAPPELL

(Invited Paper)

Abstract-A new class of computer systems has emerged which makes some insight into the origins and usages of the most common
extensive use of the fact that computers operate equally well in processing methodologies developed as a result of AI research.
either numbers o r symbols. The hest known of the systems which exploit
The current proliferation of ES’s results from the simultane-
this capability are the expert systems coming into wide use in industry.
The research from which these systems are derived is called artificial ous occurrence of several events. These include an awareness
intelligence. In spite of the popularity of systems based on this of the power and usefulness of symbolic processing. They also
technology, there is still confusion about the meaning of the terms, and include increases in the speed of processors. But perhaps most
how the technology can be effectively used. This article presents an important were drastic reductions in the cost of both process-
overview of the field, with definitions, recommendations, and an example.
ing power and memory. Increased capacity in both of these
A definition of terms is presented, followed by a brief history. Potential
roles which a symbolic processing system might play are discussed, with a elements has made the use of sophisticated graphics and a
brief summary of those which have been most successfully used. The more natural language interface practical. This has been
various methodologies in use and their respective strengths are discussed. accompanied by a reduction in cost which makes the use of
Important factors to be considered in selecting a problem are mentioned, personal workstations reasonable. Also, the availability of
followed by a discussion of the hardware and software tools available.
inexpensive expert system shells has allowed many people to
The discussion of tools includes a n evaluation of the kinds of systems for
which they seem to be best suited. The implementation effort is discussed try their hand at developing expert systems (rule-based,
in detail, because it has often been underestimated. Commitment of key mostly) on their personal computers. The capability now
personnel is a recurrent problem, and the evocation of key elements of the available in low-cost workstations has therefore accelerated
knowledge used represents an especially difficult problem. Therefore, the the spread of tools from the AI laboratory to industry and this
issues of knowledge acquisition and knowledge representation are
trend is expected to continue.
discussed in some detail. Current arguments about the best language to
use, such as the choice between Lisp and Prolog, are addressed. A A definition of the term artificial intelligence is followed
practical example is described in some detail, to illustrate how the various by a brief summary of the history of the field, including some
elements of the technology may be blended to achieve a useful autono- of the projects which defined and demonstrated the methodolo-
mous underwater vehicle. The paper closes with a discussion of current gies. Guidelines are presented which have proven useful in
research issues, including the types of machines being investigated. The
determining the utility of AI techniques in solving particular
strengths and potential of parallel architectures and neural net machines
are briefly discussed, because of their potential impact on the field of all problems. These guidelines describe the general characteris-
processing. tics of a problem which make it a good candidate for applying
Keywords: artificial intelligence, knowledge engineering, expert sys- AI to its solution.
tems. It is intended that the potential user of AI get sufficient
information from this review not only to identify the utility of
I. PURPOSE
the AI approach to his specific problems, but also to be able to

A NEW CLASS of systems has emerged which exploits the

fact that computers actually cannot differentiate between
representations of numbers and representations of symbols,
define reasonable limits for his expectations. It must be
emphasized at the beginning that the tools of AI provide a
means by which human expertise may be captured in a
and therefore have the capability to do symbol processing as machine, thus allowing it to solve problems previously solved
easily as they do number processing. In its most common only by the human. They do not allow solutions of problems
form, this type of system is called an expert system (ES). The which have never been solved before, or for which the solution
ES originated in laboratories conducting research into ways in procedures are not implied by successful human behavior.
which digital machines might be made to mimic intelligent This is the foundation for the proposed definition of artificial
human behavior. The nature of the research involved caused intelligence:
the term artificial intelligence (AI) to be applied to it and to all
of the technology developed from it. In spite of the wide use of The term artificial intelligence denotes behavior of a
the resultant technology, it is not always clear what the machine which, if a human behaves in the same way, is
specific meaning of the term is, or how problems can be considered intelligent.
identified as candidates for application of the methodology.
It is difficult to extend this definition, because the definition
The intent of this overview is to provide some definitions, plus
of what factors describe human intelligence is not clear.
One alternate definition does exist which captures the nature
Manuscript received October 30, 1987; revised February 9, 1988.
A. B. Simmons is with the Submarine Signal Division, Raytheon Company, of the work being done in the field. This was proposed in Rich
Portsmouth, RI 02871. [81], and reads as follows:
S. G . Chappell is with the Marine Systems Engineering Laboratory,
University of New Hampshire, Durham, NH 03824. Artificial Intelligence is the study of making computers do things which, at
IEEE Log Number 8821754. the moment, humans do better.

0364-9059/88/04OO-0014$01.OO 0 1988 IEEE

Authorized licensed use limited to: Bahauddin Zakariya University. Downloaded on March 11,2023 at 08:26:35 UTC from IEEE Xplore. Restrictions apply.
 SIMMONS AND CHAPPELL: ARTIFICIAL INTELLIGENCE 15

The second definition has the advantage of capturing the system will also be limited by the extent to which the selected
current preoccupation with clever use of von Neumann tools represent a “good fit” to the components of the problem
machines (given the usual image evoked by the term com- and its solution.
puter), rather than investigation of new kinds of machine, but The discussion is divided into six parts. These are
it ignores those researchers who are working on the realization
the brief history mentioned above,
of fundamentally different machines, such as the connec-
a definition of common terms,
tionist, or “neural net,” machine. Even though neural net
a discussion of possible uses: why and when,
machines are designed and built by people outside the current
an overview of available tools,
group of AI researchers, they are so closely related that they
a discussion of current implementation issues,
must be included. Therefore, the Rich definition is now too
a description of a specific application in the field of ocean
restrictive.
engineering,
These definitions contrast sharply with the goal stated in
a discussion of research issues.
Haugeland [57]:
An extensive reading list is provided. This list covers not
The fundamental goal of this research IS not merely to mimic intelligence or
only the references cited directly in the paper, but also articles
produce some clever fate. Not at all. AI wants only the genuine article:
machines with minds, in the full and literal sense. and books containing discussions of specific developments and
Very few workers in the field would agree with this some of the philosophical issues involved. For example, in the
statement, and most seem to feel that it is important that discussion some previous projects will be mentioned which
expectations be kept to a reasonably low level. It will become have been instrumental in developing and demonstrating the
clear, in the following discussion, why the goal stated by various tools now in use. These will not constitute a complete
Haugeland is not applicable to systems now being developed. summary of the work in the field of AI. However, some of the
Even so, Haugeland has presented an excellent history of the most notable are described in Waterman [86]. The list of
field and his book makes a good supplement to this discussion. references appended has therefore been made entensive
The confusion about AI may stem from the fact that, in enough for detailed examination of the background and
humans, intelligence seems to involve both an overt ability to original use for all of the topics and tools discussed, plus
solve a specific class of problems and a latent ability to others. It also allows study of some of the philosophical issues
discover solutions for a new class of problems. These two which have arisen as the use of the tools has become more
kinds of capabilities are quite distinct. Confounding them extensive.
leads to expectations that an ES will provide new solutions, References [I] through [25] provide insights into the origins
when it cannot. For example, current systems do not have the of many ideas developed in studies sponsored by government
ability to learn, in any meaningful sense, and they do not have agencies which have affected implementations. The Berliner
the ability to do induction, in the sense of creating new papers [2]-[4], [29]-[34] are important because they describe
concepts by abstracting from experience or by rearranging the evolution of the design of the current machine chess
pieces of knowledge in novel ways. Above all, that is, they do champion, a machine which has performed quite well against
not experiment. human grand masters. This design is an interesting use of
AI technology does provide a set of tools which allow some parallel processors, and uses a unique variation on the “mid
aspects of human behavior to be easily transferred to a max” search strategy for decision trees. References [l], [5],
machine, and the techniques used encourage a new kind of [6], [lo], [ll], [13], [15], [18], [20], and [25] explore issues
thought about the nature of such behaviors, because they focus in the structures of knowledge representation and control
attention on the type of knowledge involved, as well as a mechanisms. Reference [7] describes goals which were met in
plausible representation of it. They provide a framework for a distributed design used in a system developed at MITRE.
implementation of a known, but inexact, method of solving a which turned out very well (see Waterman [86]). References
problem. They do not provide the solution itself. The [8], [12], [14], [17], [22]-[24], and [67] discuss potential roles
capability provided therefore resembles most closely a new that knowledge-based systems might play. References [9] and
form of calculus, which may or may not be applicable to the [21] describe some of the problems involved in identifying the
problem at hand. It is still necessary for the user to choose a nature of the human competence being captured. Reference
specific technique, based on his understanding of the behavior [2 11, in particular, by identifying the unique capabilities of
involved and the structure of the problem. The user must vision and hearing, illustrates the difficulty of providing a
supply all the information, knowledge structures, and opera- single mechanism which mimics human capability with both
tions needed. It is therefore better to refer to the new type of senses in a single design. Reference [I] represents an
endeavor as knowledge engineering, and the systems created interesting attempt to completely formalize human reasoning
as knowledge-based systems. The resulting system will always behavior, but is extremely turgid. References [27] and [45]
be limited to the subset of human knowledge embedded in it, illustrate some of the thinking about intelligent machines
although its behavior may at times be unexpected. The scope reflected in science fiction. The publications by Borning [35],
and adequacy of the implemented system will, in any case, be [36] are extremely important, because they show how simply
limited by the insights of the people whose knowledge is being and effectively all of the systems created by humans, both
captured, and by the skill of the knowledge engineer who is concrete and abstract, can be designed, modeled, and tested in
doing the capture and transforming it to machine form. The a simple object-oriented environment (in this case, Smalltalk

Authorized licensed use limited to: Bahauddin Zakariya University. Downloaded on March 11,2023 at 08:26:35 UTC from IEEE Xplore. Restrictions apply.
 16 IEEE JOURNAL OF OCEANIC ENGINEERING, VOL. 13, NO. 2, APRIL 1988

’77). References [37]-[39], [50], [52], [71], [73], [77], [78], role of robots, to evince intelligent behavior. Later, George
[82], [85], [90], and [92] offer additional views of some of the Boole developed the Boolean algebra as a means of describing
issues involved and the approaches taken in addressing them. possible operations of the brain in a formal calculus. The early
The problems of explaining behavior by inferring the hidden period culminated with the work of Alan Turing, who
processes involved in cognition and learning are illustrated in enunciated the Turing test, and who described the design of the
1281, W I , W I , [461-[481, [521, 1531, 1561, [581, 1631, [661, Turing machine. Turing’s work provided the standard against
[77], [til], and [88]. The background of Lisp is covered in which all machines and grammars (including human gram-
[69], and the use of Lisp is covered in [85] and [94]. The mars) were evaluated for many years.
original impetus for using both Boolean logic and the binary Support for the ideas of Boole was provided in the late
form of perceptron is given in McCulloch and Pitts [70], soon 1940’s by McCulloch and Pitts [70], who discovered that some
to be republished by MIT Press in a book of selected papers brain cells operate in the binary fashion required by Boolean
being edited by Anderson and Rosenberg. The issues involved algebra. Since the same kind of logic was being used in digital
in designing a language for AI research, and in programming computers, there was an implication that the computer should
in general, are covered in [43], [57], [62], [641, 1841, and 1851. now begin to reflect more aspects of human behavior, which
Smalltalk ’80 [54], [ S I , [65] is especially important, because led to more discussion.
the object-oriented paradigm is becoming more and more In the same time frame, von Neumann defined the machine
pervasive in knowledge-based system design. The discussion architecture which bears his name and spurred the use of
of fuzzy logic in Negoita [74] is important, because it digital computers as general-purpose devices. The fact that
represents the first attempt to formalize the human capacity to programs and data could be stored in the same memory meant
work with flexible measures and with objects which may also that the machine could work on different tasks by simply
belong partly to one class and partly to another. The ease with installing the required set of instructions and data, with a
which natural language can be used when one is willing to mechanism to select the appropriate set when needed. This led
include knowledge of semantic functions is made clear in to the hope that a single type of machine could adequately
Winograd [90], the source of the “blocks world” so often solve all problems, which resulted in a focus on finding ways
cited in the literature. The issue of the Japanese lead in the to make it behave intelligently. The preoccupation with clever
field of computers based on the insights of A I is discussed in operations in the von Neumann machine introduced a bias
Feigenbaum [49]. References [68], [72], [80], and [89] which is still evident in AI research.
discuss neural net machines, the brain operations on which In the mid-l950’s, research started in earnest on ways of
they are based, and some of the design issues. Minsky and mimicking human behavior by processing images (or sym-
Papert [72] is important because the negative evaluation bols), primarily at MIT and Stanford. The research originally
contained therein halted funding for neural net machine followed two paths: development of unique hardware and
research for about 20 years. Lippman [68] is a good, short development of programming tools for von Neumann ma-
survey of the field of current machines and dominant chines.
implementations. Hardware research culminated in the perceptron, which was
To illustrate the practicality of AI technology and methods, so severely criticized in Minsky and Papert [72] that research
a detailed example is given. The example chosen shows how on what are now called neural net machines stopped for some
the various elements of the technology may be artfully blended 20 years.
to achieve a system capable of reasonable behavior. The The term artificial intelligence was coined in this time
system selected is an unmanned, untethered (i.e., fully frame to describe the research being done.
autonomous) underwater vehicle. This design is especially A belief in the efficacy of stepping through simple rules,
interesting because it makes no attempt to reflect expertise, per combined with a belief that the operations of the Turing
se. Instead it focuses on two facets of the behavior spectrum: machine accurately model human behavior, led to claims for
behavior achievable in digital computers which could not be
1) real-time reaction versus reflection before action; substantiated. The turning point in this phase seems to have
2) “reasonable” behavior versus “expert” behavior.
been the project for automatic translation of languages. This
The system is structured so that expertise can develop failed so badly that it provoked a complete cessation of funding
gradually, and can be added to the system without requiring an for A I research in Great Britain and a severe reduction in the
extensive redesign. Because of this, it represents a uniquely United States.
attractive use of all methodologies, and it is an especially good The unsatisfactory outcome of projects undertaken in the
example of the utility of the techniques of using multiple 1950’s caused the whole field of AI research to become
blackboards and using the object-oriented paradigm. relatively dormant until the mid-1960’s. Soon after its renais-
sance, impressive machine behavior was achieved, especially
11. HISTORY in problem solving and diagnostic assistance. The ability to
The research which led to the current AI technology started reflect the behavior of an expert was impressive, even though
after the Second World War, and was spurred in large part by the range of behavior was quite limited. Landmarks of this
the advent of the digital computer. However, the foundations period include:
were laid much earlier. The “computing engine” of Babbage
provoked much discussion on the potential of machines, in the development by McCarthy [69] of the LIst Processing
I

Authorized licensed use limited to: Bahauddin Zakariya University. Downloaded on March 11,2023 at 08:26:35 UTC from IEEE Xplore. Restrictions apply.
SIMMONS AND CHAPPELL: ARTIFICIAL INTELLIGENCE 17

(Lisp) language, which focuses on recursive processing talented people. The architecture used, an “inference engine”
of arbitrary lists; plus a “knowledge base,” became the prototype of the class of
the “blocks world,” from Understanding Natural systems now denoted by the term expert system. The success
Language, by Terry Winograd [9 I], which demonstrates of this work gave impetus to the application of AI technology
limited communication with a robot in natural English of all types, but especially inference engines, to all kinds of
and with automated planning; practical problems.
The Eliza program, which mimics a Rogerian psycho- Several companies appeared on the market with practical
therapist, by Weizenbaum [87]; products. Hardware for machines specializing in the Lisp
The Semantic Organization of Memory, by Quillian language (“Lisp systems”) was presented by Symbolics, Lisp
[73], which outlines a system of world linkages reflecting Machines, Inc. (LMI, now Gigamos Systems), and Xerox.
concept linkages in natural language; The OPS5 software system was made available for the VAX
the MYCIN program [42], which demonstrated the utility computers, running under Unix. Comprehensive, specialized
of rule sets and confidence factors in developing a software for knowledge engineering on Lisp systems was
diagnostic adviser for physicians. presented by Intelligenetics (now Intellicorp), Inference
Projects such as these supported the view that existing Corp., and the Carnegie Group, among others. An excellent
machines, given the right operating system and the right set of tool for use on the Macintosh (Nexpert) was presented by
rules, could accurately emulate important kinds of human Neuron Data, Inc. The field had now matured.
behavior, especially when the role of the machine was Two types of methodology have become dominant. The first
carefully selected. They also supported the view that Lisp was is concerned only with sets of rules and their activation. The
the language of choice for investigating intelligence, because second is concerned with more general problems of the
they were done in one or the other of two dialects of Lisp: the organization of knowledge, plus a need for cooperation
Interlisp dialect (BBN, Xerox, and Stanford) or the MacLisp between various sources of knowledge in a synergistic way.
dialect (MIT). The Interlisp-D dialect dominated in this work, Systems using only the rule-based approach focus on capture
because it was the dialect used at the Palo Alto Research of appropriate rules, the facts used, and proper rule execution
Center of Xerox (Xerox PARC), and because it reflected an in a limited, specific situation. These are the kinds of systems
emphasis on user interaction resulting from the work of properly referred to as expert systems, since they focus on
researchers such as Alan Kay. The importance of the Xerox capturing the insights of an expert in a small domain, and
PARC work of this period can be judged by the fact that the coding these insights. The second type of system encompasses
Macintosh computer, the Sun workstation, and the Apollo a larger domain than just facts and rules for one situation. The
workstation, as well as all comparable machines, use ideas latter type of system is better referred to as a knowledge-based
generated during this time at Xerox PARC. However, the field system.
remained relatively dormant. While expert systems have been dominant, it is also realized
These developments, coupled with the split of software that they have limitations. These limitations are being over-
activity in academe into the separate discipline of computer come in the more generally oriented knowledge engineering
science, caused a divergence of research activities on robots systems by the adoption of other methodologies. These include
and research on machine intelligence. Researchers in machine “state-transition networks” and “case grammars. The
”

intelligence (AI) have not been especially concerned with “frame” construct, a form of knowledge organization pro-
robotics, and roboticists have not generally been concerned posed in Minsky [l], has, in particular, become very impor-
with (or used) AI. Researchers in AI have been primarily tant in knowledge engineering. Thus, several very powerful
concerned with programming existing sequential machines, systems for organizing and using knowledge were beginning
accepting as given the constraints thereby incurred, because to evolve, based on a combination of all of the insights of the
software has been their primary concern. Roboticists, on the late 1960’s and early 1970’s.
other hand, have been primarily concerned with hardware The major limitation of rule-based systems stems from
useful at the end points of behavior sequence: sensing and difficulties encountered in extending or modifying rule sets.
motion. Sight, touch, “hearing,” and motion control have This is especially true when the only form allowed is the
been treated by roboticists as amenable to deterministic “conjunctive normal” form (only combinations containing
approaches (i.e., not involving heuristics). For this reason, the “and”) as is recommended in using Prolog. When the rule set
subject of robotics is not covered in this review. become relatively large, there are problems with tracing the
The beginning of the 1980’s saw a proliferation of projects effect on other rules resulting from changes in a specific rule,
based on the work done in the late 1960’s. During this time, a and with predicting the effect on existing rules resulting from
new organization became prominent: Carnegie Mellon Uni- the addition of new rules. This stems from two facts:
versity. This group developed a rule-based problem solving
1) There will implicit “or” operations because several
system called Rl/XCON. This was developed under the rules produce the same result.
auspices of DEC to provide assistance to people responsible 2 ) There tend to be some rules which trigger activity when
for translating purchase orders into collections of specific units it is inappropriate.
and cabinets, including cabling orders. The system developed
for this purpose was one of the first to be labeled an expert It is now recognized that one of the primary abilities of a
system, because it was based on the behavior of especially human is selection from a large set of rules of a small subset to

Authorized licensed use limited to: Bahauddin Zakariya University. Downloaded on March 11,2023 at 08:26:35 UTC from IEEE Xplore. Restrictions apply.
 18 IEEE JOURNAL OF OCEANIC ENGINEERING, VOL. 13, NO. 2, APRIL 1988

apply in a given situation. Recognition of the situation is experts to develop machines which can increase human
crucial in this process. It is not clear how the same kind of capability by acting as “knowledge amplifiers. Two goals ”

recognition can be easily done in a system which is built of seem to be common. One goal is the use of heuristics to do
only rules and an executive which activates a single rule for low-level processing, thus speeding up a task while freeing the
each step. The use of the frame methodology alleviates this human to do more creative work. The other goal is to improve
problem. the performance of individuals by making available the advice
The final result for the field of AI is the emergence of and guidance of an expert. Success in achieving both of these
products and applications, discussed in the remainder of the goals is a major reason for the popularity of the ES approach.
paper, most of which involve use of the ES approach. The prevalent implementation procedure requires involve-
ment of several types of people to realize the desired system. It
111. ROLESAND INTELLIGENT ACTIVITY involves the activity of the hardware experts and programmers
The definition of intelligent activity is as described above: traditionally involved in developing digital systems. It also
observing behavior in a machine which evokes the term involves a direct interface between a knowledge engineer, who
intelligent when the same behavior is observed in humans. specializes in discovering and encoding knowledge, and an
Granted that this type of behavior can be evoked from expert. The tools and the hardware to be used will generally
machines, the question becomes the extent to which the have already been selected at the time the implementation
machine can then substitute for the human in the same type of starts, so that apriori constraints are imposed by the design of
situation. the hardware and software selected. Because of these a priori
There are several roles the machine might play: constraints, care must be taken to avoid a “black hole”
situation, in which increasing effort results in decreasing
“gofer”: getting routine data, or performing simple,
effectiveness.
repetitive tasks;
A lot of human interaction is required. The need for a new
intelligent assistant or adviser, who can help keep track of
type of engineer-the knowledge engineer-results partly
information and progress, providing advice when needed;
from the fact that experts often cannot explain precisely why
replacement for the human in a specific type of activity.
they make specific decisions, so that simple interviews are an
The analysis activity involved in all of these roles includes not inadequate way of capturing the relevant cognitive process.
only awareness of relevant facts and their implications, but More often, a careful study of context and activity, using the
also the recognition process. The recognition process, in turn, skills of an objective observer, is required to discover the most
requires retrieval of a model which defines not only all of the likely triggers of relevant activity. The problem is quite
relevant facts, but also all activity, assumed behavior, and similar to the discovery process in which anthropologists,
evaluation procedures needed to plan and execute a program of linguists, experimental psychologists, and cognitive psycholo-
activity. gists (among others) are trained. This is why it is often found
The role of human replacement is often thought of in terms that people from these disciplines make better knowledge
of the speculations in science fiction stories, where the engineers than people trained in engineering or computer
machine replacement is “smarter” or “better” than the science.
human replaced (or, as hypothesized in Delaney [45], a human It is recognized that this situation is unsatisfactory, as it adds
reprogrammed in a machine language to act like a machine). a new layer of personnel to the layers that already exist, the
This possibility is not seriously considered. Replacement of number for standard systems being shown in Fig. l(a) and the
humans is considered, but in a very restricted sense, in number for knowledge-based systems being shown in Fig.
situations where the environment is very hazardous. However, l(b). Currently, the usefulness of a machine created indepen-
there is no expectation of intelligence beyond the hope of dently by hardware engineers is obtained as a result of the
asymptotically approaching human capability, and none of the activities of a programmer, who translates agreed specifica-
successful systems now in place are of this type. tions into machine behavior using a high-level language
The “gofer” type of system can be implemented on existing overlaid in the machine language. The emergence of the
machines if the range of activity is severely constrained. If knowledge-based system created a need for the specialist in
implemented on a Lisp machine, an interesting level of both discovery and programming to translate observed behav-
sophistication can be achieved, as shown in the blocks world ior into the acceptable behavior of an existing machine. The
robot of Winograd. That is, the system can have an ability to tools of the knowledge engineer are dialogue and observation
absorb new information, to answer questions, and to carry out in “representative” situations, followed by extraction of
orders, all its inputs being in normal English, and all of its sequences and situations which seem relevant. This is an
responses also being in normal English. The key point is that inexact procedure, which gives varying results. The uncertain-
the universe of activity and meaning must be severely ties involved in articulating ‘‘logical” sequences for behavior
restricted, as indicated by Winograd. Within this constraint, which is not, in itself, clearly sequential or logical creates
the machine can provide excellent support for users. This has further problems, especially when an attempt is made to merge
led to a proliferation of machines operating as an intelligent the rationalizations obtained by working with several experts.
assistant, adviser, or tutor. Therefore, the eventual goal is to let the expert have a machine
The major effort, therefore, is currently focused on devel- with which he communicates directly, and which learns
opment of systems which can use the knowledge of existing directly from him, as shown in Fig. l(c). In this final phase,

Authorized licensed use limited to: Bahauddin Zakariya University. Downloaded on March 11,2023 at 08:26:35 UTC from IEEE Xplore. Restrictions apply.
SIMMONS AND CHAPPELL: ARTIFICIAL INTELLIGENCE 19
.....................................
j Event ,
4 I
Interenrlr

ILI
I

(a) (b)
U*er/KmBrt mmage
* Rasp on8e

kfq
m m m m mnouusr source Activity

CW”ltl”@ Bq1ns.r
i Expert System i
.....................
(C) Fig. 2 . Expert system.
Fig. 1. Trend in capability of computer systems. (a) Standard computer
system. (b) Knowledge-based system. (c) Cognitive system (future).

of their declarative nature, and this has spurred the develop-

the user works alone with a self-organizing machine, teaching ment and use of the Prolog language for rule-based systems.
it directly to behave as desired. This phase requires the The semantic network structure has been most commonly
development of machines which depart radically from current used in defining taxonomies. Typically, it is shown with two
von Neumann designs, and is most likely to arrive when the kinds of link: “is-a” and “a-kind-of,’’ as shown in Fig. 3.
neural net machine technology matures and fuses with other In this structure, each higher level node is considered to be
technologies. the parent of all of the nodes below. Each lower level node is
the child of all of the higher nodes to which it is connected. As
IV. DEFINITION OF TERMS
a result of this linkage, each child node can be made to
The most common terms for the basic structures of the automatically “inherit” all of the attributes standard for its
knowledge-engineering discipline are class, with a provision to override the standard inheritance
1) expert system with unique attributes, if desired. In addition, each node can
2) semantic network have unique attributes applicable only to itself and its children.
3) frames The instance shown is a very simple case. The links can be
bidirectional, with the added label “is-part-of,” if desired. It
4) blackboard.
is also possible to have other types of label, such as “has-a” or
The single ES is envisioned as consisting of two parts: an “can,” and to have as many types in a single network as
inference engine and a knowledge source, as shown in Fig. 2. desired. Each node can belong to as many sets as desired, and
The inputs to the engine are events. These may be the can thus “inherit” characteristics from as many “parents” as
statement of a new fact or the negation of a known fact or desired. These links are a natural extension of the “attribute
belief. These events may be spontaneous, as in the case of a list” construct in Lisp. The number and types are limited only
sensor, or may result from a search for possibilities, as in the by the number of pointers which can be stored. This type of
case of medical diagnosis. The knowledge source consists of construction is used in the search trees (see Charniak [42],
those rules needed to react properly to any combination of Rich [77], or Winston [92]). It also serves as the basis of
values of the facts within the purview of the system. When a object-oriented programming (see Stefik and Bobrow [84]),
new fact is discovered, or when an event necessitating where each of the nodes represents an object with knowledge
assertion of a new fact occurs, the inference engine examines and the ability to process messages. There is a problem in
all of the relevant rules to determine whether or not a response applying these concepts to natural language, because every
is appropriate. These rules all have one of two forms: “if A, node contains many other nodes within itself when defined as a
then B,” or “C is implied by A and B.” In the form “if A then part of the language (see Quillian, in Minsky [73]).
B” the first part (left-hand side), is called the premise, and the The most common use for the search tree format is in
second part is the “production.” In the form “C is implied by representing the activity required to achieve a specific goal,
A and B,” the first part is the fact to be asserted wherever both given a set of facts. This occurs in diagnostics, identification,
A and B are true. The second form reflects an approach based and games. In these cases, it can be assumed that all of the
on symbolic logic. facts are independent and can be linked together in all possible
The production specified by any rule of the first form is combinations. Achievement of a desired goal then involves
executed when the rule is activated by a specific combination tracing the paths through the tree in the most efficient and
of true and false values of facts. In the event that the premises economical manner. The two simplest approaches involve the
of more than one rule are satisfied simultaneously, a “conflict use of “breadth-first” or “depth-first’’ searching. In the case
resolution” procedure is invoked to select one production of breadth-first, all of the nodes at the current level (called a
only. ply) are examined before stepping to the next level (this
The meaning of the second form for a rule is often taken to corresponds to the kind of operation followed by a simple rule-
be “A and B must be done before C occurs.” based system such as OPS5). In the case of depth-first, a single
The definitions and operations used in both kinds of rule path through all of the nodes succeeding the current node is
make this kind of system realizable in symbolic logic because traced to the “bottom” of the tree, before path branching from

Authorized licensed use limited to: Bahauddin Zakariya University. Downloaded on March 11,2023 at 08:26:35 UTC from IEEE Xplore. Restrictions apply.
 20 IEEE JOURNAL OF OCEANIC ENGINEERING, VOL. 13. NO. 2, APRIL 1988

ANIIUL

/ / .\IS-A
left wall g I center w a l l h right wall

POODLE AFGHAN
1 IS-A
e
floor
SPOT
Fig. 3. A basic semantic net.

Frame Itern: Defining Characteristlcs:

the right-hand neighbor is traced. In either case, the search Wall:
Left : Edges ”a“, “d” & “g”;
normally stops when one of the terminal “leaves” is the Corners “ag” & “gd”;
desired goal. The complexity of the operations involved is etc.; no inconsistent comblnatlons
typified by the problem of playing chess. Exhaustive search of Center: Edges e, b, g & h,
all possible combinations of plays is impractical, even if a Corners “ e g d , “efh”, “bag” & “bch“,
etc.; no l n c o n s l s t e n t combinations
supercomputer is used. The basic problem with this approach 0
is that the search must operate in an informed way; if the
problem is combinatorial, explosion is to be avoided. In the

t
Contents:
METADENDRAL system for identifying organic compounds Table: - Deflned by Dining Room Frame
(see Waterman [86]), this was done by using the “cause- Chalrs:
effect” constraints imposed by knowledge available from etc.
additional sources to constrain acceptable combinations of (b)
nodes and branches in the tree, drastically pruning it, and thus
Fig. 4. Basic frame concept. (a) Frame elements, room. (b) Dining room-
drastically limiting the size of the search space. In the case of basic expectations to which defaults are assigned.
chess, the tree must be pruned by a combination of search
through several layers and assignment of values to paths which
I
include moves by the opponent. The most successful example
plate
of this is the system of Berliner [29]-[34], which uses a system
of cooperating experts coordinated by a strategy manager. The
ratings applied by Berliner to dynamically choose the best
path, thus dynamically pruning the tree, follow the same
philosophy as the “confidence-factor’’ approach used in
chaird o r
systems for diagnostics (see the MYCIN system (Davis and
Lmat [ U ] ) )The. use of causal reasoning to constrain the
search is covered very nicely in Davis [25]. The element patio
common to all of these approaches is to direct the search along
promising branches in the paths through the tree. The general
name for this kind of operation, where only a few candidate
--r
branches are selected for further examination, is beam-
directed search. This type of search produces a more efficient door
and satisfying result than the simple stepping procedures used Fig. 5. A dining room frame.
in either the depth-first or the breadth-first operations.
The other major construct in common use is the “frame.” plausible. It also attempts to account for the invariance of
Fig. 4 shows the sample of this construct as defined by perception in the presence of translation and rotation. In terms
Minsky. In this case the frame will attempt to capture the of reaction to a situation, it postulates a selectivity operation
features characterizing a “dining room,” one example of based on the perception which allows humans to bring to bear
which is given in Fig. 5 . The data structure representing the a small subset of knowledge and rules appropriateto the task at
significant elements is shown in Fig. 6. Notice that frames can hand. In this view, there exists some “typical” case which
be nested within frames. This construct was defined to somehow constrains the processing and interpretation of facts
describe a metalevel construct which might help to account for to that set which matches expectations.
some human behavior in handling complex situations and The data structure defining a frame consists of a list of
scenes. It attempts to incorporate the use of partial knowledge attributes (or pointers to sets of attributes), procedures, and
by postulating the existence of “defaults,” where the default is rules which describe the “typical” case of the object or
some concatenation of fundamental characteristics (called situation being analyzed. In principle, all the attributes listed
significant features) which allow a wide range of objects to be are unique to the object, and constitute the necessary and

Authorized licensed use limited to: Bahauddin Zakariya University. Downloaded on March 11,2023 at 08:26:35 UTC from IEEE Xplore. Restrictions apply.
 SIMMONS AND CHAPPELL: ARTIFICIAL INTELLIGENCE 21

DINING ROOM FRAME for the frame approach. However, an attempt to use this kind
of structure to organize knowledge in the standard system can
FURNITURE always help clarify details of the design and operation, and can
Table Massive, Leaves (Optional) be very helpful in constraining expectations.
Visual Cues Table Furnishlngs, Dining
Chairs Back, Seat (Solid, 15in f r o m floor), Legs,
The blackboard organization of knowledge-based systems is
A r m r e s t s (Optional), important because it provides a natural motivation for the use
Visual Cues Back, Part or All of Seat, of distributed processing. This type of system is fully
0-4 Legs, no mismatches discussed in Nii [76]. The blackboard concept involves the use
Sideboard
FURNl SHl NGS
of specialized knowledge sources which communicate through
Rug(s) or Carpet a common data base. In the original Hearsay and HASP
Draperies implementations, shown in Figs. 7 and 8, the knowledge
DECORATIONS sources were rule sets, each using the same data base as the
Wall Plates, Pictures, Sculptures, Tapestry, e t c
others. Fig. 7 shows the basic system structure, and Fig. 8
Ceiling Textured w i t h Sculptures (Optional),
Painted w i t h Pictures (Optional) shows the flow of information between levels of expertise.
LI GHTl NG These instances also used a centralized executive, which
Large Ceiling Lamp (Single, o r Cluster) operated the knowledge sources according to global rules of
Wall Lamps (Optional)
priority and the specific type of knowledge needed next. In
Candles (Optional)
ACTORS both systems, the key operation was generation of a current
Diners I through 5 best hypothesis, following the functional flow shown in Fig. 9.
Servants 0 through 2 This provided the filter through which events were passed, and
SCRl PTS
could be confirmed or denied by future events.
Breakfast
Lunch The Hearsay system attempted to analyze speech. The
Supper or Dinner knowledge sources and their organization reflect this. They
are organized in a hierarchical fashion, with a sound analyzer
Fig. 6 . Dining room data frame. at the lowest level and a phrasal analyzer at the highest level.
The levels shown reflect the hierarchy used in linguistic
sufficient set to characterize all members of the set (as in studies. Decoding sound to determine its meaning involves
“dining room” and “dining room table” in the example). In decoding acoustic waves to locate segments of the phonetic
addition, the attribute list contains some characterization of alphabet, using whatever values a preprocessor finds for
plausible boundary conditions, such as the character of selected characteristics. These can be frequency-based, as in
connecting objects (one dining room door normally opens onto the case of “formants” (characteristic frequency combinations
a kitchen, not a beach, as in the Minsky example). That is, the for specific phonemes), or time-based, as in the case of typical
frame accurately characterizes a prototype, so that any item of transitions between phonemes. Above this, specific combina-
the generic type is instantly recognized, no matter what its tions of features are joined into typical phonemic segments of
variations, within some kind of intuitive limits. However, words. Combinations of phonemic candidates are joined to
there are severe problems in using this method of description form candidate words, creating candidate syllables along the
in a computer. These are: way. The candidate syllables are joined at the lexical level by
using knowledge from the phrasal level, which has access to
1) It is difficult to get agreement on the precise items
semantic information. This is shown in Fig. 8(a) by the
occurring in a “typical” case.
addition of the conceptual level. This provides the hypotheses
2) It is difficult to scope the range of all the items in the
that are to be tested to finalize the decision as to what was said.
general category “plausible.”
It is important to note that a final selection must be delayed
3) It is difficult to encapsulate the ability of the human to
until the activity of all of the knowledge sources has been
categorize a large variety of different items as plausible
completed. That is, the combinations at the lower levels
alternatives in a specific case.
cannot be reduced to one (1) until a “consistent” (or
4) It is difficult to emulate the ability to make correct
“plausible”) phrase has been detected. Therefore, feedback
identification, based on having available only parts of the
must be used to assist in the final selection at all levels, and
items.
may involve reexamination of raw data to see if a new
5) It is difficult to provide, without a large amount of
combination can be justified that better supports a probable
processing, the ability to recognize objects regardless of
form for the utterance. The system must therefore suspend
size and orientation.
final judgment until the knowledge sources have reduced the
These problems are intriguing, because humans seem to solve possible form to only one choice by looping several times
them so simply and so well. Wall decorations could be art or through the facts, with modified hypotheses to be tested.
plates or mirrors. The chairs need only be partly visible. The The same design philosophy is reflected in the HASP/SIAP
identity of objects is not lost when passing by them. system, shown in Fig. 8(b). In this case, the levels in the
This kind of generalized capability is beginning to be hierarchy are determined by the content of the information
realized in neural net machines, and that type of machine may handled. In this design, the lowest level encompasses signal
be the only practical way to cover all the possibilities needed detection and signal processing, a functional activity similar to

Authorized licensed use limited to: Bahauddin Zakariya University. Downloaded on March 11,2023 at 08:26:35 UTC from IEEE Xplore. Restrictions apply.
 22 IEEE JOURNAL OF OCEANIC ENGINEERING, VOL. 13, NO. 2, APRIL 1988

tact
Knouledge implication Knovledge
Level 1 0 source 1

Knouledge
Level 2 a e fact
Knoulsdge
source 2

tact Fig. 9. HASP design. One global data structure contained a hypothesis about
1.Plicntirm Knavlbdpe the situation. After each cycle of the plan, hypothesize, validate predict
Knouledge 0 source 3 operations, the hypothesis changed states. These states were called the
Level 3 preliminary hypothesis, the working hypothesis, the current best hypothe-
sis, and the predicted hypothesis.

validated by information supplied at the fleet level. The fleet

- 1 I level handles the most general information, and provides
hypothesis formation information from outside sources. Incor-
Fig. 7. Basic blackboard design.
poration of other sources at the highest level allows anticipa-
tion of the appearance of a source, for example, and this
LEVELS : ComCT*oNS ' KNOWLEDGE SOURCES. allows processing to be sensitized to the expected emissions.
This latter facet of the behavior emphasizes the fact that the
effectiveness of the system is strongly dependent on the
accuracy and flexibility of the models which derive the
hypotheses transferred from a higher level to a lower level to
provide pruning, or to drive a more detailed analysis of data
previously connected.
The cyclic operation of HASP is shown in more detail in
Fig. 9. In a given time interval, the operation starts with test of
a prediction. This is the link between the "predicted hypothe-
sis" block and the "plan" block. The plan block encompasses
all of the analysis activity needed to verify or discard the
selected hypothesis. This results in the generation of a refined
or preliminary hypothesis concerning the types of sources.
These are validated or discarded on the basis of comparison
with data provided by simulation or processing of real data.
The result of the data comparison test is then used to generate a
new current best hypothesis as to the sources(s). The current
best hypothesis is then used to predict what will happen in the
next unit of time. If all goes well, the "predict-and-test''
operation causes convergence on a single answer for the
identity of any source of emitted sound which has been
detected.
The major design issue in the Hearsay and HASP systems
concerns activation of specific knowledge sources when an
event occurs. The executive program, which controls activa-
the acoustic processing activity in Hearsay. These functions tion of each source in turn, must have accurate knowledge of
are shown separately as the segments, lines, harmonic sets, which sources are interested in what data, so that the
and sources levels. Within the levels, the same type of appropriate knowledge source is scheduled when a fact is
selection process occurs as in the Hearsay system. The inserted. The same executive must have heuristics for dis-
segment layer extracts groups of acoustic characteristics which patching knowledge sources in such a say that the more
might correspond to lines from emissions from typical important events are processed first, so as to minimize the
sources. Potential lines are grouped into families at the level of number of times the loop must be traversed and to minimize
harmonic sets. The harmonic sets are associated with specific the number of retractions required because of the fact that
sources at the level of platforms. All of this is coordinated and conclusions generated by one knowledge source require that

Authorized licensed use limited to: Bahauddin Zakariya University. Downloaded on March 11,2023 at 08:26:35 UTC from IEEE Xplore. Restrictions apply.
 SIMMONS AND CHAPPELL: ARTIFICIAL INTELLIGENCE 23

previous conclusions by other knowledge sources be revised. similar way, the ability to make the scarce resource of
In the examples cited by Nii, the types of knowledge used expertise more widely available is an excellent motivation for
are visualized as being organized hierarchically. That is, development of systems to assist in medical diagnosis or in
semantics represents a higher level of knowledge than phonet- making financial decisions. Extending the reach of the
ics, and fleet knowledge is at a higher level than line expertise of a good teacher is a type of situation that occurs in
knowledge. In any case, the organization takes into account teaching, and this has given impetus to the use of AI
the interaction between the types of knowledge such that techniques for computer-aided instruction.
knowledge at the higher level can be used for two purposes: to Beyond these simple cases, however, the decision is not so
predict future events and to confirm current conclusions. The clear-cut. Often, it seems to be “because we want to do AI.”
key element, however, is the independence of the knowledge This is not considered the best approach. A possible set of
sources, not rank. Each source is a specialist in a narrow guidelines which provide an alternative is as follows:
domain, much like human specialists. This isolation into 1) The process being investigated is performed in a
specialties makes this architecture ideally suited for implemen- heuristic manner by a human. That is, the solution
tation by means of the object-oriented paradigm. depends primarily on judgment, not on calculations.
It is not necessary to use the type of implementation 2) The process requires searching large data spaces, thus
described by Nii, however. In general, each of the knowledge being indicative of a complex combinatorial problem.
sources could be a different machine (specialized, if desired), There is no unique or “best” solution, but a set of
and communication could be by broadcast over a bus (or acceptable solutions. In addition, judgment is required to
buses), with all relevant data captured and stored locally. That determine which solution, within the set of solutions,
is, the system could be implemented in a way completely in should be used. This may involve generating and testing
accord with the rules of the object-oriented paradigm, as multiple hypotheses in order to compensate for missing
shown in the example discussed in Section VIII. Each level or uncertain data.
may also have buses to the other levels. Message traffic can be It is desired to separate the data base from the executable
determined by the kind of data needed by other machines. In a code, so that the data base can be easily modified as a
dynamic system, this can change as a result of each machine result of corrections or additions.
registering to receive specific types of data from a distribution It is desired to capture, in permanent form, the reasoning
center, thus permitting an interesting adaptability to be and/or reaction process involved in arriving at a solu-
achieved. One feature common to all of the systems so far tion, so that reliable solutions can be generated for
implemented is adaptation of the data analysis to expectations. situations or cases which are different from previous
The system maintains a set of hypotheses, based on its world situations or cases.
model, and uses these to focus the data-analysis effort on The process being investigated requires the modeling of
expected events. Actual events or facts are then used by all human behavior.
knowledge sources to revise plans and hypotheses. The only The development process requires that several candidate
difficulty in this type of approach is control of the number of designs be quickly outlined for comparative purposes
cycles needed for convergence. The looping operation must be (rapid prototyping) .
executed once for each new or revised fact. The problem is in It is desired to have easy and natural communication,
trying to determine the maximum number of loops needed for with adaptability, between the user and the system.
each contribution of each knowledge source to circulate
It must be remembered that the von Neumann machine is
through the system and be digested by all knowledge sources
most useful for high-speed, repetitive operations. It is there-
such that no new individual contribution triggers further
fore an ideal source of assistance in tasks which require this
changes. No method is yet known for predicting convergence
ability. The need to account for details which otherwise might
or stability in such a system.
be overlooked or the need for rapid and tireless application of
V. PROBLEM SELECTION routine analysis to large masses of data (a Turing machine, but
with the use of heuristics rather than algorithms alone)
When it is decided that AI techniques should be investi-
indicates a task which is a prime candidate for the use of AI
gated, the first rule of problem selection is to choose a problem
technology.
that appears to be too simple. It will become clear, as the
implementation proceeds, that it is sufficiently difficult. It will VI. AVAILABLE TOOLS
be made especially clear when the prototype system is The following sections discuss the software and hardware
demonstrated to a group of experts. Many comments will be found to be most useful in developing knowledge-based
offered about all of the things it does not do. Within this systems. It is important to note that none of the tools described
general constraint of staying with “simple” problems, some below are directly usable in real-time control problems. In
guidelines are described below. those situations where the problem involves replicating expert
The most obvious situation in which AI techniques apply is behavior in a real-time problem, such as control of a chemical
in the capture of the thinking of a recognized expert, especially or petroleum plant, an additional issue of real-time data
in the case where the expert is about to be lost. The collection and response must be addressed. This requires the
opportunity to capture such expertise in permanent form in a use of tools and technology being developed by GigaMOS,
processor so that it is usable by others is compelling. In a Inc. (the PICON system) and Gensym, Inc. (the G2 system).

Authorized licensed use limited to: Bahauddin Zakariya University. Downloaded on March 11,2023 at 08:26:35 UTC from IEEE Xplore. Restrictions apply.
 24 IEEE JOURNAL OF OCEANIC ENGINEERING, VOL. 13, NO. 2, APRIL 1988

These tools and systems are unique and cannot be integrated

with the tools discussed below.
REASONING PROCESS:
When a fact i s asserted:
A . Software Tools 1. I f the asserted fact is already true, nothing happens.
The tools available can be roughly divided into the two 2. If the fact is new infur&ht?lt will become part o f the
classes cited above: those primarily concerned with rules and knowledge base.
3. Rules in the speclfled rule class that have a fact in thelr
those concerned with general knowledge units, including Premise matching the fact are ’tickled’ (identified).
rules. The former is exemplified by OPS5 (Carnegie Mellon 4. ‘Tickled’ rules are Lested individually. If all of the facts
University) and Nexpert (Neuron Data, Inc.). The latter is in the premise of a rule are true. it is added to the
forward-chaining agenda (list o f active rules).
exemplified by KEE (Knowledge Engineering Environment by 5. As each rule is removed from the forward-chaining agenda
Intellicorp) and ART (Automated Reasoning Tool by Infer- and applied. i t s assertion (or assertions) become part
o f the knowledge base.
ence Corp.) 6. I f the rsserfiun o f an applied rule causes the premise of
Rule-based systems focus on selection of rules whose another rule to be completely satlsfled. forward chaining
conditions are satisfied by the current set of facts, and continues on the new assertion.

“firing” new rules when a new fact appears, or is “asserted.” Fig. 10. Forward-chaining, chaining procedure.
The Nexpert system is somewhat more sophisticated than a
simple rule system manager, in that it provides a lot of
graphics and a very friendly user interface, coupled with very
effective support for tracing rule dependencies. The Nexpert Forward Chaining can be invoked by using the keyword ‘assert‘
system also allows rules to be grouped into sets which are used with a rule class:

only in a specific situation, thus adding a type of frame (ASSERT ‘well-formed-formula ‘ruleclass)
capability. t causes reasoning to start
in this class
All of the available tools include some form of inference basic assertion: ’add
engine. In the case of such general-purpose systems as For Example:
ART, KEE, and Nexpert, the environment provided also (ASSERT
contains tools for handling frames (as unique objects), ‘(the entree o f dinner is lasagna)
‘choose.wine.rules).
semantic nets (as dependency lists and/or a hierarchy of where the fact being asserted matches a factive phrase in the
inherited traits), and procedures. Among the extra tools premise of one or more rules (‘the x of y Is 2.1.
provided is a means for retracting all previous decisions Fig. 11. Forward-chaining, activation.
affected by a change in knowledge (called a truth maintenance
system). The KEE and ART systems also provide a means for
experimenting with facts and situations to determine the
choose .wine .rule immediately affected :
viability of alternative interpretations of facts, or the effect of
adding new rules, without disturbing the current system; this is (if (the entree of- is ?&.d
called an alternate worlds investigation tool. In any case, the (the color of * l s 1 T
asagna
general-purpose systems represent an extension of the simple (hh is i n WINES)
rule-based system and can therefore provide examples of hearty-burgundy (deduced from next clause)
(thecolor of 3xine is-
operation for all paradigms and knowledge representation hearty-burgundy red (becaus;. i t is the only one
techniques. i n WINES with red; ‘hidden’ rule)
then do
In a any rule-based system, rule activation may proceed in a (a potential.wine o f ? m d is b i n e )
“forward-chaining’’ manner as shown in Figs. 10-13. dlnner hearty-burgundy
The example chosen is taken from the tutorials used for the Fig. 12. Forward-chaining, initial step.
KEE system. There are variations in terminology and syntax
among the various systems, but the basic procedure is the
same in all systems. The forward-chaining procedure is started
by the change in state of some fact in the system, or by the
insertion of a new fact. In the case of KEE, both cases are this -triggers-:
covered by the “assertion” operation. As a result of the (If (a potentlal.wlne o f IdlnaerIs 2rf.lne)
dinner hearty-burgundy
assertion, the procedure described in Fig. 10 is executed.
(the status of- is onhafull
Steps 2 through 5 represent the sequence of activity repre-
sented by the term forward-chaining. Step 3 is taken because
the fact represents new information, and the independent ‘status’ = attribute o f object
nature of the rules in the rule base requires examination of all -
hearty-burgundy’. and
‘on.hand one of i t s legal values
rules. Each rule is examined to see if its premise (the “if” then do
(the wine of Zdinwc Is Mne)
part) is satisfied as a result of the existence of the fact. If so, Note the implied ’meta-rule’ that the color of the wine and the entree
step 5 is taken, and any new facts resulting from execution of should match
the operation(s) specified in the assertion clause (the “then Fig. 13. Forward-chaining, final step.
do” part) are added to the knowledge base as new facts. This

Authorized licensed use limited to: Bahauddin Zakariya University. Downloaded on March 11,2023 at 08:26:35 UTC from IEEE Xplore. Restrictions apply.
 SIMMONS AND CHAPPELL: ARTIFICIAL INTELLIGENCE 25

causes step 3 to be repeated, and steps 4 through 6 as well, if

appropriate. This execution loop continues until no more rules
are active (the agenda of rules to test is empty). In the example
shown in Figs. 11-13, it is desired to find out what wine is
available which is suitable for consumption with a dinner Use Query wlth a rule class named:
entrke of lasagna. The rules capture the heuristics such as the
(query ‘well-formed-formula ‘ruleclass)
general rule that a strongly flavored red wine should be
consumed with a strongly flavored red entrke such as lasagna, as in:

and that the desired type of wine should be among those on (query Ythe wine of dinner is ?what) ’choose.wine.rules)
hand. These heuristics are captured in a knowledge unit called
Note:
the choose.wine.rules. In this case, the assertion phrase
matches a factive phrase in the rule shown in Fig. 12: “the The query pattern must match the 8SSefhOn (COnClUSlOn)
of one or more rules in the rule base to iniliate backward
entree of ?meal is ?entre,” where the ? indicates a variable chaining.
which may have any one of a group of values. In this case,
Fig. 14. Backward-chaining, initiation.
“dinner” and “lasagna” are acceptable. The system then
proceeds to test the rest of the premise, to see if it is totally
satisfied. In this case, lasagna is found to have the color red,
and “hearty burgundy” is found in the group of objects called Rule System Goals:
WINES. The action which accomplishes this assertion has Each query pattern posed establishes a goal, i . e.. a problem to
be solved.
been labeled a hidden rule, because it is implicit in system The query that initiates the backward chaining may be viewed as
operation; the system automatically examined the object class the fop-fewfgoa4 as in OPS5.
When a rule is tried. as a part o f a possible path to a goal, lhat
WINES and returned all objects with the attribute “red.” The rule‘s premise orpremises also become goals.
result of this sequence is that the fact in the assertion clause of A goal can be resolved as being true or non-true in one of three
ways:
this rule can be asserted, so that a possible wine selection is
By flndlng the facts needed for the goal In a relevant
indicated. This assertion causes the rule in Fig. 13 to be tested. knowledge base, or in a collection of general facts, or
The sequence of tests on the phrases in this rule indicates that By deriving the relevant facts. using a rule, or
there is some hearty burgundy on hand, so it is selected, thus By asking the user If the needed value Is valid
completing the search, since there is no factive phrase of the Fig. 15. Backward-chaining, procedure.
form “the wine of ?dinner is ?wine” in any of the remaining
rules.
Alternatively, the rule-based system may operate in a
“backward-chaining’’ manner, as shown in Figs. 14-17. In (query (the wine o f dinner is A@d)
’choose.wine.rules)
this type of operation, a goal is specified, and an attempt to
% (if (a potential.wine of-ls 3
justify it is made by checking to see if supporting facts can be dinner
found. That is, an attempt is made to derive the conclusion (the status of is on.hand)
from existing facts by tracing backwards through the premises then do
of rules. Necessary facts may come from interaction with the (the wine o f ? d i n n r r i s
user or they may be derived by noting other rules whose
“firing” would supply missing facts. The example shown in
dinner
I
the figures is also taken from the KEE system. In this system,
the goal of the search is defined by the keyword query. The Fig. 16. Backward-chaining, sequence of steps
“well-formed-formula’’ (wff) is any instance of a legal type of
factive phrase used in the conclusion part of rules (called the
assertion in KEE). The procedure invoked as a result of the
query is described in Fig. 15. The fact that the wff in the query %
. (If (the entree of-Is 2W.c~)
matches the wff in the conclusion of some rule causes the

1
dinner lasagna
(the color o f m i s Ir9cnd
potential satisfaction of each of the wff‘s in the premise of that lasagna red
rule to also become goals, secondary to the goal established by (the color of 2alae is 2wkxl
the query, which is the “top-level” goal. (In OPS5 and related hem-fy burqtmdy red

systems, the nomenclature is “goal” versus “subgoals”.) In hearty burgundy

the backward-search procedure, in contrast to the forward- then do
(a potential.wine o f ‘ A
lahnec is 23rLw)

I_
search procedure, the secondary goals can be achieved by the dinner ??
derivation process (“pattern matching” in Prolog) or by
asking the user. In the example, the wff matches the assertion %%%THEANSWER: HEARTY BURGUNDY
clause of the rule which asserted the wine to be consumed with Fig. 17. Backward-chaining, the answer.
the entree for dinner. It is found that the second phrase of the
premise is satisfied, and the first phrase establishes a new
goal. This goal is satisfied because of the fact that the factive

Authorized licensed use limited to: Bahauddin Zakariya University. Downloaded on March 11,2023 at 08:26:35 UTC from IEEE Xplore. Restrictions apply.
 26 IEEE JOURNAL OF OCEANIC ENGINEERING, VOL. 13, NO. 2, APRIL 1988

phrase which generated it is matched by the assertion of the larger variety of systems. The KEE paradigm is especially
rule shown in Fig. 17. Setting out to satisfy all of the goals interesting because it encapsulates all items in frames, so that
created by the wff‘s in this premise results in the discovery that they can be tested as independent objects, and then disciplines
there exists a wine called hearty burgundy for which all all interactions to follow the object-oriented programming
necessary conditions are satisfied, which allows the desired paradigm. In the object-oriented paradigm, every object
answer to be generated. belongs to one or more classes. Each object inherits qualities,
Most tools provide a means for doing either of the rules, and procedures from the classes to which it belongs, as
operations illustrated above to some degree, but the sophistica- well as having its own unique qualities, rules, and procedures.
tion and ease of use varies widely. The more sophisticated, To this extent, it reflects the semantic net concept. However,
such as Nexpert, ART, and KEE, provide automatic switching each object interacts with all other objects by sending and
between the search mechanisms. This is accomplished in receiving messages-there is no such thing as a program or
several ways. It can be done by causing the system to switch in subroutine call in the normal sense of these terms. In this
accordance with an agenda, by establishment of new worlds respect, it follows the paradigm used in implementing
which use modified rules and search procedures, or by noting Smalltalk (copyright by Xerox) and LOOPS (copyright by
that assertion of an additional fact would cause a particular Xerox), the original object-oriented programming systems.
rule to be enabled. It can also be done by other techniques. The Flavors system of Symbolics is available for this type of
such as querying the user. All of these features make the rule work on Symbolics machines. Other suppliers are using
system more flexible, and are especially useful during the CommonLOOPS, the CommonLisp version of LOOPS. Thus,
design phase. They are obtained at the cost of increased CommonLOOPS is likely to become dominant, and should be
bookkeeping, to provide a trace of all linkages, both forward considered when it is desired to program a unique system
and backward, between rules in each rule set. Therefore, they which requires portability.
require larger memories and cannot be easily implemented on Current recommendations for selection of an appropriate
the PC type of machine. However, provision of these features environment for the development of knowledge-based systems
also allows automatic notice of the impact of changes in tend to favor KEE for systems in which evaluation of
individual rules or in rule grouping. This is a feature not situations, prediction of behavior, and execution of procedures
available in the systems available on the small PC’s or in the are primary considerations. The ART system tends to be
public version of OPS5. Sophistication is also added in the favored in cases where pattern-matching and search operations
larger systems by provision for the extensive use of confidence are primary. These recommendations are based on the fact that
factors and “fuzzy” logic. The use of confidence factors KEE provides a somewhat more sophisticated capability in the
allows path activation (by selection of the next rule to fire) to use of frames and semantic networks and has a reasonable rule
proceed in accordance with a numerical estimate of the degree system with truth maintenance and an extensive graphics
of certainty in assessing a cause and effect relation between capability. This seems to result from its origin as a tool for
facts. It also allows the final conclusion to be limited to a set of geneticists. The ART system is oriented somewhat more to the
possibilities with relative likelihoods. The fuzzy logic concept programmer type of user. It provides an excellent pattern-
recognizes that specific “things” may belong to more than one matching capability and very good tracing of reasoning
class, inheriting from all parents. It also provides a way to sequences. The ART system is the source of the worlds
account for the fact that there may be multiple causes for concept, which allows testing of the effects of rule changes
events and the fact that there exist “uncertainty intervals” in without disturbing the current system. It reflects its origin as
many processes. These concepts have been treated in both an extension of the rule system in OPS5, and is well suited to
mathematical and common sense ways, as described by situations in which symbolic logic is used heavily, as in
Negoita [70]. There are currently no firm guidelines to guide a searching operations.
choice of which approach is appropriate in any given case. For situations in which the domain is smaller and the
The more powerful tools provide both powerful graphics activity reasonably restricted, the Nexpert system is recom-
support and a large set of tools for development. They include mended. It has a well-designed system for grouping rules by
systems for automatic truth maintenance, so that any and all the situation in which they are used, an excellent system for
conclusions or actions voided by a new fact can be retracted. graphically browsing through a network of interdependent
The KEE and ART systems, for example, provide a tool for rules, and an extremely simple user interface. However, it
experimentation, called worlds, which allows investigation of requires the use of a graphically oriented system, with bit-
the effects of changing facts, rules, or procedures, as noted mapped graphics, such as the Macintosh.
above. This construct allows extension of the rule sets by The cost of the tools ranges from nothing for a public
merger of worlds, where the system supplies feasibility version of OPS5 for the VAX to about $50 OOO for the first
indications. Because of their sophistication, these tools are copy of either ART or KEE, with Nexpert somewhere in the
recommended for use in cases where a reasonably large middle (about $7000). The relative power of the tools is
system, subject to evolution and requiring interaction between roughly correlated with their cost, as indicated above.
several sources of knowledge, is to be implemented. Beyond the tools themselves, it is important to consider the
Both the KEE and ART systems provide knowledge issue of language used and the ability of the system environ-
structures based on the frame concept, and both have been ment to support languages other than that used by the tool. The
rewritten in CommonLisp to make them portable across a language used by the tool is important because it always

Authorized licensed use limited to: Bahauddin Zakariya University. Downloaded on March 11,2023 at 08:26:35 UTC from IEEE Xplore. Restrictions apply.
 SIMMONS AND CHAPPELL: ARTIFICIAL INTELLIGENCE 27

becomes necessary to write new procedures to augment those software, with the release of the Coral Corporation’s version
supplied. For almost all tools, this means that the knowledge of CommonLisp.
engineer must be fluent in Lisp. Support of other languages is Both the workstation and the personal computer alternative
important, so that existing code can be executed during allow early infusion of AI technology into a commercial
operation of the system, for the sake of reducing the cost of concern at low cost, with low risk.
implementation. The latter situation occurs when the knowl-
edge-based system is to be used as an adjunct of an existing VI1. IMPLEMENTATION
ISSUES
system, and the existing system already has extensive code for
Given that a knowledge-based system is judged useful, and
modeling and simulation. Easy access to the existing code
that the role of the system has been decided, there remain three
becomes mandatory, in this situation, and is becoming a
general areas of concern in implementing the system. These
common feature in all systems. However, the ease with which
are
other software can be “spliced in” dynamically varies
considerably. One example of a good mechanism is provided knowledge acquisition
by the operation of Nexpert on the VAX. This tool makes knowledge representation
good use of the “virtual image” feature of VMS to allow system architecture.
programs in Oracle, C, or any other language to interact
These topics are discussed separately in the following para-
dynamically in a manner transparent to the user.
graphs. Because they are somewhat interrelated, the discus-
sions overlap.
B. Hardware Tools
A . Knowledge Acquisition
The hardware systems available to run the major software
systems include dedicated Lisp machines and standard ma- The knowledge acquisition task can consume 50 percent to
chines running a dialect of Lisp or c. 80 percent of ‘the total effort of implementing a system,
The dedicated machines belong to two classes: workstation depending on the amount and type of knowledge to be elicited.
and mainframe. The workstation machine includes the Xerox There are two sources of problems in this process. First, as
and TI families. The mainframe type of machine is built by noted above, the expert often cannot articulate reasons for his
Symbolics, Inc. The cost varies from about $25 OOO for the behavior in the form of standard rules. Second, the expert
Xerox machines to about $100 OOO for the largest Symbolics typically does not have time available for extensive interviews,
machine. The throughput varies roughly as the cost of the especially because his expertise is in short supply. The first
machine, but the concept behind lower cost machines is that problem evidences itself when the expert gives a set of rules
high throughput may not be necessary. In fact, the extensive and then observes the results. These are usually unsatisfactory
on-line memory and large mass memories now available on on the first pass, and the rationalizations originally given can
the smaller machines make them quite adequate for many even be counterproductive. For example, the expert may
tasks. This fact, coupled with the fact that all of the machines, visualize using facts stated in the input data, but they turn out
regardless of price, support only a single user effectively, to be implicit, rather than explicit. That is, the expert may do
makes the smaller machines an attractive alternative. Use of some “preprocessing” of which he is unaware. The quality of
the TI machine is made even more attractive by the possibility the rules given verbally is thus about the same as the quality of
of using the TI “Lisp machine on a chip” as a part of a those given by a native speaker of a language when he attempts
delivery vehicle. TI now supplies a Lisp processor in a 1/ to justify the forms used in his language. The fact is that they
2ATR configuration, suitable for use on military aircraft, so are just what is done by custom, not because they are
that there is now credibility to the concept of a deliverable necessary. But the native speaker feels that a more universal
system for use in difficult environments. and logical reason should be given, when asked why.
Most of the mainframe and workstation machines provide Therefore, as a general rule, satisfactory completion of the
the capability of running the same knowledge-engineering knowledge-acquisition task will require careful observation by
software, but at less speed in the workstation. The problem is the knowledge engineer of the behavior of the expert in
that the stack processing and garbage collection functions different situations. This must result in careful notations of
required for the Lisp language create a heavy load for existing changes in the environment to supplement answers to ques-
machine designs. tions asked of the expert. The discovery procedure includes
An even cheaper low-end capability is becoming available the use of videotapes plus notes on index cards, the technique
on personal machines. It is now possible to transport sophisti- used by linguists in field work in an unknown language. The
cated software such as Nexpert to these machines, so that the discovery process must then be followed by constant observa-
newer versions of the IBM PC and the Macintosh can now be tion of the emerging expert system by the expert.
used as AI workstations. The IBM PC can, in fact, be In general, the knowledge acquisition process therefore has
modified to become a very powerful workstation in Com- anthropological aspects and linguistic aspects. Above all, it
monlisp, so long as graphics are not needed, by adding a involves the general problem of trying to duplicate human
special board (the Hummingboard), supplied by Gold Hill behavior without knowledge of what significant features the
Computer. The Macintosh currently can be used for the human responds to and without conducting a sensitivity
Nexpert software, and may soon be usable with other analysis to determine them.

Authorized licensed use limited to: Bahauddin Zakariya University. Downloaded on March 11,2023 at 08:26:35 UTC from IEEE Xplore. Restrictions apply.
 28 IEEE JOURNAL OF OCEANIC ENGINEERING, VOL. 13, NO. 2, APRIL 1988

B. Internal Knowledge Representation The frame structure is especially attractive because it

The problem in analyzing the knowledge of the expert can focuses on the context in which behavior occurs. This means
be restated as resulting from the fact that the knowledge that a new frame can be chosen when the context changes,
engineer has to codify the behavior of a black box whose inner allowing a rather natural change in the set of rules used. It is
workings are not visible. This problem is made worse by the also useful because it is natural to include the actors creating
fact that the activity to be codified is not verbal in nature, and the situation, with their expected behavior. Standard behavior
cannot be easily translated into verbal form-sometimes it can be included in the form of a script, and this can be used to
cannot be translated at all. Successful analysis involves guide the selection of rules and hypotheses. Scripts, in turn,
knowledge of significant features whose nature and identity can be derived from standard missions or goals, thus creating a
are unknown at the start, and which are amenable to rigorous strong and well-motivated causal link between all of the
description only by execution of a sensitivity study too elements of the system, as well as allowing deduction of facts
complex to undertake in the time usually available. Beyond and actions not explicitly stated (see Schank [78]). When all of
this, there are issues of the form and process involved in the elements of the frame structure are considered together,
cognition and concept formation by which the mental models the frame is also very helpful in indicating the actions required
are composed. The nature of these facets of knowledge to retract all the effects of a conclusion when events require a
represents an issue which has not yet been resolved. The fact change in beliefs.
of the existence of items such as an idealized conceptual model The problem with using frames is that, often, no real
(or prototype) seems obvious, but their structure ind rationale situation is exactly the same as the “typical” situation, so a lot
have yet to be adequately defined. Even if they could be of thought must be given to the variation allowed within a
defined, it is possible that they may not be accurately realized given situation before a change in situation must be postulated.
in existing machines, because of the profound differences The deviations from “typical” or “normal” which must be
between the machinery of the brain, as compared to existing accommodated in frame-based systems suggest that the use of
computers. fuzzy logic may be required, but the kinds of variation
Unless this is recognized, implementation of a satisfactory allowed, and their extent, can be difficult to enumerate.
system will be difficult. It is clear that the knowledge The tree structure, on the other hand, can be used to show
acquisition process transcends simple logical operations and the specific nature of relationships between elements of the
requires much more effort than simple interviews. The knowledge. It can also be used to search for the best action to
problem is complicated by the oft-imposed constraint that the take next, or to help decipher a sequence of words, as in
model developed for the knowledge of the expert must be augmented transition networks, where each word restricts the
composed only of the elements of traditional logic. set of following words which are plausible. However, the
evidence that humans do not search sequentially through a set
C. External Knowledge Representation of rules must be kept in mind. Evidence now available points
There is a significant problem in choosing the framework in to a single-step process for recognition of a stimulus and
which the knowledge of an expert is to be inserted. This selection of a response. Humans seem to operate sequentially
problem arises from the fact that the result must be usable in a only in speaking and moving. All other operations appear to be
sequential machine which has only two dimensions of inter- done in parallel, in very few steps. The point is that encoding
connection, which has very few connections between compo- knowledge in a tree form is reasonable for showing semantic
nents, and which is restricted to sequential operation in any relationships, but it is a difficult form for defining search or
one processor, so that human activity cannot be easily classification operations, because it introduces artificial, linear
described, if at all, in the simple system of linguistic and logic constraints. As a result, encompassing a large set of cases is
structures available. In addition, humans operate with partial difficult, and changes are difficult. This means that refinement
data and with degrees of uncertainty not easily captured in of knowledge is difficult, so that the sequence of changes
either the language of the machine or its logic. It is generally inherent in the process of progressing from novice to expert
true, therefore, that a mixture of methodologies will be needed are difficult to enable. Therefore, tree representations of
to provide a reasonable emulation of human behavior. This knowledge, including state diagrams, are a very weak model
means that the system will make use of all three of the basic to use in attempting to emulate human behavior. Attempts to
structures described above. In all cases, however, it must be use this type of structure in a von Neuman machine graphically
remembered that success depends on the extent to which the illustrate the limitation of the computer, as compared to the
domain of the problem can be limited and the behavior of the brain.
system constrained to a small, well-defined set of possibilities. Rule sets are satisfying for modeling simple decision
The frame structure is useful in formulating typical situa- processes. To allow rapid development and test, plus easy
tions and exposing the assumptions associated with them. This maintenance, the number of variables must be small, and the
is important in systems which must advise on appropriate combinations few. Rule-based systems are easy to build-rules
behavior. For this purpose, the fact that the frames can include can be generated at the rate of several hundred per month. It
predicted behavior and methods for generating hypotheses soon becomes clear, though, that interactions between the
makes them a powerful tool. This is over and above the ability rules are too complex to easily trace and control without
to minimize the number of rules which must be checked when special tools, and may be difficult even then. It will always be
any designated type of event occurs. necessary to introduce control methods, such as frames, and

Authorized licensed use limited to: Bahauddin Zakariya University. Downloaded on March 11,2023 at 08:26:35 UTC from IEEE Xplore. Restrictions apply.
 SIMMONS AND CHAPPELL: ARTIFICIAL INTELLIGENCE 29

tracing methods, such as semantic networks, to avoid disas- the alternatives provided realistically mimic human activ-
trous side effects when changing the rules. Therefore, use of ity;
rules alone should be undertaken only for very small prob- the implementation chosen allows evolution in a simple
lems . and straightforward way.
There is evidence that the ease of system implementation
All of these features of the design create a unique concinnity.
varies in proportion to the complexity of the tool. That is, the
The most important of the tools used are the multiple
designer can quickly build a system which uses only a set of
blackboard and the object-oriented paradigm. This extensive
unordered rules. Building a system which uses sets of related
use of the latest developments represents another important
rules with controls for selection of sets or rules by means of
reason for discussing it, instead of previous projects.
semantic relationships takes longer. Expanding the knowledge
In addition to the good selection of tools, however, the
structure to include frames causes the implementation to take
system design is organized to reflect a reasonable model of
longer still. It has also been found, however, that a large staff
human activity. The reason for this is that the design is
is required to maintain a large, unordered rule set, because of
hierarchical, with separate levels handling different parts of
the extensive debugging required whenever changes are made.
the time domain, so that “reflex” actions are independent of
It is likely that most problems are sufficiently complex to
“planning” actions. Further, each knowledge source and
require the extra effort involved in using the complex tools
execution module operates as a coequal of all other modules,
because most problems are not as simple as originally
in a collegiate type of organization. This minimizes the
assumed. This is the reason for the rule that the selected
capability required of the coordinator and facilitates imple-
problem should appear to be overly simple. Things are rarely
mentation in other architectures, such as parallel processors.
as simple to do as they seem to be.
The design is in the process of changing, to incorporate
insights gained from experiments conducted with a prototype
D. Language
model. Therefore, this description represents a snapshot of the
There are cases in which the system is to be entirely design in the form it is currently taking.
constructed by original programming, not by using existing The multiple blackboard part of the design is based on the
tools. In this case, the user can select the language felt to be dual-blackboard approach defined by Barbara Hayes-Roth
most appropriate for the kind of problem and/or the approach w11.
selected. When this occurs, there are several choices for doing
A . Missions
object-oriented programming, and one major alternative for
. rule-oriented programming. The EAVE vehicle is both unmanned and untethered, so that
The major choices for doing object-oriented programming it is completely autonomous. It will have the capability to
are Smalltalk-80 and CommonLOOPS. Smalltalk-80 has the explore or map an underwater area independently, given only
advantage that it is available on the Macintosh and the a statement of the goals of the mission in general terms. An
Tektronix workstation. CommonLOOPS has the advantage example of the kind of terrain it might be asked to operate in is
that it is the portable paradigm for workstations and main- shown in Fig. 18. The assigned mission might be one of the
frames. All of the various paradigms, including Flavors, are following:
discussed in Stefik, [80]. exploration: map bottom and location of objects;
The major alternative for rule-based systems is Prolog. This salvage: locate and pick up object;
language is favored for declarative types of programs because military action: attack, decoy, or neutralize.
it has a rigorous formal logic, good pattern matching, and
good resolution rules. Examples of problems in which Prolog An unmanned vehicle allows operation in hostile environ-
has been useful include parsers for computer languages and ments. The use of an untethered vehicle provides greater
relational data base languages. The basic criterion indicative freedom of action, and avoids potential tangling problems
of its value is that the knowledge is wholly factive in nature. when obstacle avoidance is required. Total autonomy is
In general, however, the tools created in Lisp have broader preferred because of the limitations on the rate at which
applicability, because different kinds of knowledge can be information can be transferred in the absence of a tether. Other
modeled and coordinated. Furthermore, the graphics and user complications avoided by the untethered design include
interface tools are more extensive and more mature. weight, maintenance, and the problem of control of payout and
payin when maneuvering.
VIII. A PRACTICAL
EXAMPLE In any case, there is an element of uncertainty in execution
of the mission which prevents the planning of a precise
The Experimental Autonomous Underwater Vehicle
program of actions. The instructions for the mission can only
(EAVE), in development at the University of New Hampshire
be considered to constitute a “wish list” of desired activity,
Marine Systems Engineering Laboratory, will be used as an
with the understanding that variations will be made as the
example of the practical use of the ideas in AI. There are
mission proceeds. These variations must include the possibil-
several reasons for this selection:
ity that all of the desired tasks cannot be performed, and the
it involves real-time operations; mission instructions must reflect this, as described below.
the design blends all major tools and methodologies; Therefore, there must exist a reasonable level of “intelli-
the multiple levels used reflect human mental structures; gence” on board the vehicle.

Authorized licensed use limited to: Bahauddin Zakariya University. Downloaded on March 11,2023 at 08:26:35 UTC from IEEE Xplore. Restrictions apply.
 30 IEEE JOURNAL OF OCEANIC ENGINEERING, VOL. 13, NO. 2, APRIL 1988

Fig. 18. To plan a mission completely, with detailed instruc-

tions covering all contingencies, prior to vehicle launch is
impractical, as in the case of tethered vehicles. Therefore, as
the mission proceeds, the path followed must bend and fold as
required to reflect the actual environment. The possibilities are
defined to be either “preplanned, with no variations allowed”
or “tentative plan, variations accepted. ” The latter approach
is chosen, and the planning is done in such a way as to
minimize commitment on any part of the plan as long as
possible, thus incorporating the “least-commitment strategy”

Fig. 18.
--
Typical mission area.
for the path selection operations. The basic constraint adopted
is that variations should maintain mission integrity, except for
well-defined contingencies.
The practicality of this approach also depends on the degree
The blackboard design approach is desired because of the to which obstacles and events can be grouped into simple
domain neutrality it allows. Domain independence is achieved classes of ‘‘typical cases. Currently, the grouping envisioned
”

by separating the domain-specific knowledge from the me- for obstacles is “wall,” “edge” (or “tube”), and “block”
chanics of applying the knowledge. Domain-specific knowl- (can include “sphere”). The set of events which may require
edge is encapsulated in the knowledge source, and domain- mission abort (or may prevent achievement of all desired
specific data are posted in the blackboard, for observation by goals) includes such things as “fuel exhaustion” and “leak in
all concerned, in any form the designer wishes to use. enclosure.” TO allow such occurrences to be handled in a
Therefore, while designing the system, the designer is free to reasonable way, a suitable repertoire is embedded at each level
implement as many knowledge sources and actions as are of activity, as described below. None of this represents a
required to solve his problem or to allow changes to be procedural difference from a standard computer program, in
incorporated. The posting technique used may consist of a execution. It does differ, however, in the manner in which the
broadcast (as on a bus) or activating a knowledge source as an steps of the execution are activated and revised. It is possible
object, with a pointer to a buffer. In addition, the blackboard at any time prior to actual execution of a step to revise or
paradigm encourages the exploitation of any modularity eliminate it. This allows any future commitment to be changed
inherent in the design. This applies to both the procedural as needed, including further examination in a given area,
knowledge embedded in the knowledge sources and the data reordering activity, or eliminating some tasks. In effect, a part
posted on the blackboard. Further, the modularity used is also of the program is being written as the mission progresses. In
flexible, so that the size of the modules can be adapted to any engineering terms, the designs may be compared as “under-
desired level of granularity, representing any hierarchical constrained’ ’ versus “overconstrained, ” where the standard
pattern desired. algorithmic approach represents the overconstrained mission
execution.
B. Design Requirements The following types of activity are anticipated:
For an unmanned, untethered vehicle to be practical, the
prior to launch: task assignment and preliminary mission
question of the model for its intelligent behavior must be
plan;
addressed. In the standard expert system approach, the model
during mission:
is a recognized human expert. In the case of exploration by
quick reaction to obstacles,
unmanned vehicles, there is no expert in the use of such
planning of existing subtasks,
vehicles, so the art must be learned as experience is gained.
specification and planning for new subtasks (i.e.,
Furthermore, it is not clear what kind of expertise in
examination of an object of interest),
exploration should be captured, given that the vehicle should
mission revision.
behave like a human carrying out the same kind of mission.
Therefore, no attempt is made to embed “expertise,” in the To encompass these types of activity, a hierarchy of three
sense normally used. Instead, the focus is on embedding levels of processing is established, as shown in Fig. 19. These
“reasonable” patterns of behavior, in the sense that they three levels are determined by the time domain in which the
reflect the activity of an intelligent human. In other words, an activity must occur.
expert in exploration is not needed for many missions; the The types of activity required are exemplified by the cycle
need is for an intelligent anticipation of contingencies, of search activities shown in Fig. 20. This figure presents a set
combined with a reasonable repertoire of responses. This of tasks representative of the type of reconnaissance mission
approach is reflected in the design. The multiple blackboard which might be assigned to the vehicle. In this figure, the
and the encapsulation of individual knowledge sources as mission consists of two search operations, indicated by the
separate objects are used specifically to allow the flexibility rectilinear patterns, and a group of points to visit, enclosed in
needed to add the results of learning, as well as the need for the large circle. The supervisor module inserted the way points
different kinds of vehicles in different environments. for the search patterns, and a human specified the way points
The need for flexibility is reflected in the contours shown in in the visitation group. The dashed lines represent paths

Authorized licensed use limited to: Bahauddin Zakariya University. Downloaded on March 11,2023 at 08:26:35 UTC from IEEE Xplore. Restrictions apply.
 SIMMONS AND CHAPPELL: ARTIFICIAL INTELLIGENCE 31

Vehicle for obstructions. Each stereotype evokes a standard response,

Analysis/ Long-Term Planning & Support
Functions based on classification by the sensor. In the event that
Reflection Reflection
something occurs which is on an emergency list or an alert list,
mission the vehicle can pause and await further instructions.
The middle level provides planning for the next sequence of
steps, with the ability to revise paths as needed, based on the
Situation
Assessment Environment tentative plan and events occurring during the mission. Partial
Level
plans for different tasks are finalized and down-loaded to the
Lbstrect,on I Medium-Term Planning
& Execution
I
System
reflex level as needed. This includes reaction to obstacles,
where the reaction is to await instructions (in the event that a
(lcamrd res-srs) system problem occurs). This level thus fits together the pieces
Assessment Guidance
of the plan as the mission unfolds, under the constraint that
lO<~T<IOOs sufficient energy remains. If a problem occurs, the top level is
notified, and a more drastic revision is undertaken at that
Short-Term Action &
Reaction
level.
The top level generates the initial plan and monitors activity
Manager Manager during the mission. The function performed during the
(quick rcastirn -&Is) execution phase consists of a continuous comparison of the
Reflexes
dT< 10s current situation with the planned situation at that point.
Discrepancies are evaluated to determine if major revisions
Environment are required. If so, a new master plan is created and down-
loaded to the middle level for implementation. Revision
Fig. 19. Levels of activity
actions can also be triggered by events logged at the lower
levels, such as fuel consumption. This activity can be
described as a constant redetermination of the feasibility of the
.IC
Start v......-.-...-..... remaining part of the mission, given the current data.
An example of how the planning cycle works is shown in
Fig. 20. The mission depicted in that figure contains three
tasks:
visit five points in a specified area;
conduct a survey on the area covered by survey A;
conduct a survey on the area covered by survey B.
f
It is clear that planning an “optimum” path is impractical,
especially in view of the uncertainties implied by the nature of
the mission (“survey”), plus possible vehicle problems.
Therefore, any “reasonable” solution is accepted, where the
term reasonable includes such things as fuel reserve and a
nominal allowance for possible problems. However, to further
ensure the execution of a useful mission, even if the initial plan
cannot be completed, each of the requested tasks is given a
priority, and the higher priority task is completed first, where
6 Finish Survey- A.
practical. This operates in the other direction, also. If
sufficient energy reserve remains and a desired visitation point
is near, it will be incorporated, even if it is not in the plan. The
allowance for problems can be made reasonably conservative
Fig. 20. Typical mission.
by assuming a “worst-case” event, such as a “wall-like”
obstruction which might require a large path change. Instruc-
specified by the supervisor module. The description which
tions to examine certain classes of objects can be executed
follows is abstracted from Chappell [41].
dynamically by changing the local path. The philosophy and
At the lowest level, the quick reaction time required makes
operation of this type of flexible control are covered in detail
the operations comparable to the reflex actions of humans.
in the Hayes-Roth paper.
Operations at this level thus have the nature of stimulus/
The major elements of the design can be summarized as
response activity. The time period for plan elements, when
follows:
loaded, takes into account the response rate of the vehicle, the
“worst” type of obstacle, and sensor capability. This level Exploit simple, general models of objects and slopes and
reports status and unplanned events to the upper levels. In responses.
carrying out their normal activity, the system elements at this Exploit limits of maneuver capability as useful constraints
level make use of stereotypes such as the wall/edge distinction for planning.

Authorized licensed use limited to: Bahauddin Zakariya University. Downloaded on March 11,2023 at 08:26:35 UTC from IEEE Xplore. Restrictions apply.
 32 IEEE JOURNAL OF OCEANIC ENGINEERING, VOL. 13, NO. 2, APRIL 1988

Employ beam-directed search, where the beam is com- discussion will describe the difference between the current
posed of typical situations and appropriate responses. design, which is a standard Hayes-Roth implementation, and
Use least-commitment strategy for execution, postponing the future design. However, it is important to note that the new
development of detailed plans until needed. implementation is designed to provide independence from
Plan in detail only for “reasonable reaction period” specific hardware architectures.
determined by maneuver envelope and typical obstruc- I ) Supervisor Subsystem Architecture: The basic archi-
tions. tecture of the supervisor component is shown in Fig. 2 l . It
Discard irreversible steps. consists of four modules: two blackboards and two sets of
Treat start of new operation as new mission (compart- knowledge sources. Each blackboard acts as a means of
mentalize, so that each task is a separate object). communication between those knowledge sources which
Distribute the expertise between levels, so that each level handle the types of data normally appearing on it as indicated.
is self-sufficient in its own domain. However, interaction between events in the domain area and
data in the control area is allowed, as indicated by the diagonal
The design incorporates ideas described in detail in Minsky
lines crossing over between the knowledge sources and the
[74] and in Agha [ 2 6 ] .Specifically, the concept of “cooperat-
opposite blackboard.
ing agents” defined by Minsky and the concept of “actors”
The underlying structure of the blackboard architecture is
defined by Agha have strongly influenced the implementation
very simple. It consists of three components: a structured data
of the blackboard architecture selected as the basic structure.
space called a blackboard, a set of reaction rules or procedures
C. Design called knowledge sources, and a scheduling mechanism for
The basic operational elements shown in Fig. 19 are as controlling the knowledge sources. The structured data space
follows: may be real or it may be virtual, depending on the implementa-
tion. In any case, the resemblance to a normal blackboard
temporal hierarchy: similar response-time activities in results from the fact that it serves as a central repository for
same level; information potentially useful to several knowledge sources;
physical separation of levels operating at different rates;
i.e., the information is made generally available, instead of
a ‘‘loop’’ structure for each level, providing semiautono-
being addressed to specific modules. The dynamics of the
mous operation. basic blackboard system are also rather simple. The scheduler
The top level is concerned with strategy and mission executes a three-step control loop:
control. It generates the initial plan, using some mixture of an
determining which knowledge should be trig-
optimization heuristic and priorities of required tasks. All
gered in response to the posting of data on the
plans are considered to be nominal, and alternatives are
blackboard;
specified, especially where segments may have to be deleted 2) selecting knowledge for execution;
or modified. However, most of the operations are done prior of the infer-
3) actual execution by the knowledge
to the start of the mission. During the mission, activity is ence operation.
restricted to monitoring reports, and mission planning only
becomes active when the designated repertoire of contingency Each knowledge source potentially posts new data as a
reactions is not adequate. Infrequent activity allows operation result of its execution, thus triggering a new cycle. This is the
at a slower rate than the lower level. sequence used in the blackboard systems described by Nii.
The middle level is concerned with implementation in detail However, the standard approach requires that the scheduler be
of the current plan, with reasonable alternatives in the event a “superexpert” who knows everything in every rule used by
that any predefined type of object or obstacle appears, in each knowledge source, plus the value of that knowledge. This
which case it adopts a predefined modification of the plan, complication can be avoided in two ways: by having each
with a notice to the top level. It is concerned with the impact of knowledge source examine a new fact in parallel with all other
the current activity on the nominal plans for the next activity knowledge sources, using it when it is applicable, or by
with detailed planning for the next activity, and execution of a allowing each knowledge source to operate once per new fact.
repertoire of contingency plans, such as modifying start and In either case, the result of the operation of the knowledge
end points or deleting segments. sources will be the posting of one or more new facts or desired
The middle level thus operates at a slower rate than the actions. The coordination module then arbitrates between
bottom level, but faster than the top level. The bottom level is solutions with a set of rules supplied for this purpose. The new
concerned with the real-time steering of the vehicle, and with approach makes use of the fact that the purpose of the
rapid response to objects or obstacles within the constraints of blackboard is communication, and communication can be done
the available avoidance procedures. It therefore operates at the in several ways. The first is the broadcast of data over a
rate required to process and react immediately to data communications medium, such as a bus. The second is the
collected by the sensors provided on the vehicle. scheduling of each knowledge source, with a notice of the new
The analysis/planning loop of the situation assessment and data as its input. The first approach is preferred, because it
supervisor modules is typical of the design at all levels, and allows all knowledge sources to be physically located in
has been more fully implemented in the experimental model. separate processors. In fact, however, a mixture of the two
Therefore, the supervisor level will be discussed in detail. The will probably be necessary in any reasonably compact system,

Authorized licensed use limited to: Bahauddin Zakariya University. Downloaded on March 11,2023 at 08:26:35 UTC from IEEE Xplore. Restrictions apply.
 SIMMONS AND CHAPPELL: ARTIFICIAL INTELLIGENCE 33

Control Ex ilion
B :tboard BI board
64 microprocessors, each of which acts as a knowledge source
EvenL Mission
for a particular square, plus a Sun workstation, which contains
1o 9 9 m Analysis the strategy expert and acts as the interface with the player.
- Pwth
Design
The 64 micros act as a collegium, in which each knowledge
Problem source has equal value. The description of each move is
Analysis broadcast on the bus to all other knowledge sources. Those
- Procedure

-
Selection
which are affected by the move examine its implications, in
Agenda Segment terms of threat value or advantage, and each recommends an
Processlng Plan
appropriate response. The strategist compares the threat
evaluations and suggested responses and selects the move to be
made. The strategist, of course, has several options, including
Control Route-Planning
employing a new tactic. It may, for example, adopt a risky
tactic, in an attempt to gain a significant advantage if the risky
move succeeds. This illustrates that a focus on the function
performed by the “blackboard” type of architecture, rather
than the details of a specific method, allows a very wide
variety of implementations to be visualized. A mixture of the
Coordination Specialists
Berliner approach and the Hayes-Roth approach forms the
basis for the EAVE design.
Modularity is a basic feature, owing to the concentration on
Fig. 21. Blackboard control architecture isolation of the knowledge sources in separate modules
because of the independence of the knowledge encapsulated in
so that some knowledge sources will have to share a CPU each one. The current design fully exploits this fact in a way
resource. In either case, however, the knowledge required in that allows the implementation to exactly follow the “object-
the coordination module is reduced, and the operation resem- oriented” paradigm. Each knowledge source is treated as a
bles the collegiate system used by Berliner [32] in his chess- separate object, whose internal operations are unknown. For
playing machine. That is, all knowledge sources are equal in each object, the external knowledge about it consists of where
their priority of firing, but the productions requested by each it is (its address), what messages it accepts, and what messages
are not equal, if they conflict. it sends. All operations are conducted by passing messages to
The fact that these variations are possible testifies to the designated objects, which can be located anywhere. This
flexibility of the blackboard architecture, and justifies its completes the isolation required to allow a parallel computer
selection as a significant advance in the field of AI. Additional architecture to be used.
advantages accruing from the use of the blackboard are as The object-oriented design which allows the implementation
follows: to be flexible also makes the design extremely adaptable. New
it is domain-independent; knowledge sources can be easily added, and the implementa-
it can be implementation-independent; tion used for a given knowledge source can be changed
it provides a mechanism for superimposing a learning without requiring redesign of other knowledge sources or of
ability; the coordination module. Thus, it perfectly complements the
it uses modularity as a basic feature; blackboard architecture.
it is easily adapted; The standard algorithmic view would use the knowledge to
it encourages a view of the problem different from the define branches at specific points in a procedure. After
algorithmic view. commitments have been made, recovery from the prescribed
course, to account for new information combined with
Independence from the domain is achieved by the fact that the reprogramming of the system, would be impossible unless all
nature of the knowledge sources does not affect the basis possibilities at all points had been predefined. The modular
design. It only affects the items posted for general consump- approach and the cooperative activity avoid this problem. The
tion, and the types of activity triggered by the activity of the approach desired in the fact of uncertainty is the linear
knowledge sources. programming approach discussed in Charniak [42]. Many of
Implementation independence results from the fact that the the design issues involved, including synchronization of the
communication role of the blackboard can be realized in many knowledge sources, reflect the problems for concurrent
ways, and the machines in which the knowledge sources are programming discussed in Holt, Lazowska, Graham, and
embedded can be completely different in nature. In fact, it is Scott [64]*
possible to include in the network machines as diverse as The dual blackboard philosophy (cooperating knowledge
neural net machines, to enhance the learning and recognition sources) is used at each of the EAVE operating levels shown in
capability without requiring any changes in the design. Fig. 19. One is the blackboard on which information relative
A good example of the way that a common bus may be used to control of the mission-the control blackboard. The other
to achieve the communication function of the blackboard is has information pertinent to execution posted on it-the
given by the Berliner chess machine. This machine consists of execution blackboard. The posting of new data on a black-

Authorized licensed use limited to: Bahauddin Zakariya University. Downloaded on March 11,2023 at 08:26:35 UTC from IEEE Xplore. Restrictions apply.
 34 IEEE JOURNAL OF OCEANIC ENGINEERING, VOL. 13, NO. 2, APRIL 1988

board, either a new fact or a modification of an existing their order, and the path between them. Their specialties
solution, is called an event. As the figure shows, knowledge therefore involve the domain and the vehicle.
sources may post to either blackboard, and they monitor both Within the control group of knowledge sources, three types
blackboards. are identified:
The descriptions which follow are based on the design
strategy specialists,
described by Chappell [41], with some modification to reflect
policy specialists,
the improved design being implemented at this time.
coordination specialists.
The control blackboard contains data associated with the
selection of strategy, the allocation of resources, and the The strategy specialist knowledge sources are concerned
coordination of knowledge sources. It is primarily used by the with planning the vehicle path so as to satisfy some heuristics
control knowledge sources. This blackboard also records the for optimum use of the resources. This might include
occurrence of internal and external events which are used to minimum energy consumption, maximizing dwell time in
monitor the progress of the mission. The data posted here are specific areas, or modifying activity around an object of
divided into the three types shown. These are interest.
The policy specialists are concerned with the ranking of
event logging, alternatives and the selection of the subset to be used by the
problem analysis, execution specialists on this mission. These specialists are
agenda processing. concerned with ordering the list of pending activities and
arbitrating between conflicting solutions and/or conflicting
Two types of events are logged: exogenous and endoge-
requests for access to resources. Included in this group are
nous. Exogenous events are events occurring outside the
specialists on procedures to use in executing search patterns
vehicle. Endogenous events are events occurring inside the
and in taking evasive action, as well as in reacting to
vehicle. At the supervisor level, the exogenous events logged
emergencies and/or recovering from disturbances.
are the input of mission request prior to the mission. Problem
The coordination specialists are concerned with notification
analysis data include information about the number and types
of each knowledge source of the occurrence of a new event,
of activity required, distances, priorities, etc. Agenda process-
with posting of proposed solutions or activities resulting from
ing refers to the development of a list of pending execution
knowledge source activity, and with maintenance of a history
steps to allow all knowledge sources to examine the latest
of knowledge source activation. These modules note the
information and post the implications, plus the information
occurrence of the event, distribute the information to con-
required to permit evaluation and arbitration.
cerned parties, and record all responses.
The execution blackboard contains data concerning the
The route-planning knowledge sources are divided into two
execution of the mission under constraints imposed by the
types:
control knowledge sources. This includes planning of pieces of
the path (“segments” grouped as tasks), as well as the order in domain specialists,
which various parts of the mission are carried out. It is also vehicle specialists.
used to communicate information about vehicle activities.
Four types of data are identified, as shown. These are The domain specialists are concerned with the detailed
planning of the path. They include knowledge sources which
mission analysis, set the starting and ending points for each segment, locate the
path design, tasks and procedures needed, expand specific types of seg-
procedure selection, ments into patterns, close up any gaps in the path, and
segment plan. otherwise refine the design. Two types of segment requiring
specific patterns are identified: the survey type and the point-
Mission analysis data consist of information pertaining to group type. The first represents an activity over an area; the
segment identification, segment size, type of activity per second represents linkage of points. These knowledge sources
segment, priority, etc. Path design information concerns the also generate commands required to vary the activity of the
start and endpoint data, maneuvers, etc. Procedure selection sensors in the event that different activities are required at
information concerns the algorithm to be used to command the different times, such as initiation of a special search pattern for
execution of the desired patterns, and also to estimate the time the survey activity.
and energy required. Segment plan information is composed The vehicle specialists group consists of knowledge sources
of the command sequence and timing data compiled from the which represent expertise in vehicle capability and perform-
planning activity. ance and which supply information needed to specify each
The control knowledge sources are specialists in selecting maneuver in detail and to assess the budgetary implications of
the strategy to be used to complete the mission to the extent the plans in terms of total energy capacity.
allowed by the current situation. They also specialize in Thus, the domain specialists are concerned with the details
communication with, and coordination of, other knowledge of the geometry of the mission and the procedures used to
sources. generate specific instructions, whereas the vehicle specialists
The route-planning knowledge sources are concerned with are concerned with the state of the vehicle and the capability of
executing the current plan by implementing a path consistent the vehicle to maneuver as desired.
with it, in terms of the number of segments and types required, The basis input to the supervisor module is a specification

Authorized licensed use limited to: Bahauddin Zakariya University. Downloaded on March 11,2023 at 08:26:35 UTC from IEEE Xplore. Restrictions apply.
 SIMMONS AND CHAPPELL: ARTIFICIAL INTELLIGENCE 35

for a mission, provided by the user. The output of the module adapted for many other purposes. It differs significantly from
is a set of directions for the areas to be examined, the actions the standard blackboard control system described by Hayes-
in each area, the order in which they are to be done, and the Roth and incorporated in the experimental vehicle in the
alternatives in the event that certain combinations of events following ways:
occur. This output is supplied to the next level. The supervisor
The knowledge sources are autonomous and operate in a
module is relatively inactive during activity at the lower
collegial fashion, with the arbitrator being the first among
levels.
equals.
The basic sequence of operations is as follows: All events
Events are effectively, if not actually, broadcast either to
are noticed by the accept-input knowledge source, which all knowledge sources or to the knowledge sources
distributes a notice of the event to all other knowledge sources.
registered to receive the type of event posted.
The other knowledge sources acknowledge receipt, and those
Arbitration takes place after each knowledge source has
who are affected by the event post a notice of further finished operating and not before any knowledge source
information to follow. The affected knowledge sources now
has operated.
become active. A list is made of the knowledge sources now
There is a provision for the standard real-time require-
active, and planning activity is suspended until all inputs have
ments for interrupts, and for priorities to be attached to all
been received. The set of responses is examined, and a single
events, including requests and alerts.
response is selected, in accordance with the heuristics for
The system is event-driven, provision being made to
priority, to be distributed as the next event of interest. This
handle both endogenous and exogenous events.
sequence continues until no new facts or productions are
posted. In this procedure, all knowledge sources are treated as Object functions and their operations strongly resemble the
equals, but arbitration is done between conflicting outputs. actions of the “society of cooperating agents” postulated in
The same type of organization is used at all levels, with Minsky [74].
variations in the details covered in the knowledge sources. For Current implementations of the blackboard system reflect
example, the tentative plan is accepted by the second level as a the programming problems associated with use of the von
set of requests, with estimated time and energy utilization. The Neumann architecture. To minimize conflict over use of a
plan is then examined as a group of segments, where the single CPU, there is a “superman” type of expert. This
separable parts, such as survey A, survey B, and point-group, superexpert combines rules of importance with his detailed
are examined as independent tasks. The result of this knowledge of all of the rules in all of the knowledge sources to
examination is generation of a detailed sequence of commands select one knowledge source at a time for operation. The
for the next segment, plus a new set of estimates for time and invocation of knowledge sources is treated another case of the
energy consumption, as required. The actual time and energy standard procedure all in typed languages, with the facts used
usage is monitored continually, so that plans for both the by the knowledge sources treated as variables. This means that
current and the next segment can be modified as needed. The the scheduler must know where and how all facts are used
knowledge sources available also provide the knowledge before the operation starts. This introduces a lot of redundancy
needed to initiate special sensor activity, such as special and makes the system both rigid and fragile. The rigidity is
transmission or search patterns to help collect information because change becomes more difficult; the fragility because
about the nature of an object or obstruction. all possible combinations must be anticipated in the rules of the
The lowest level contains the highest concentration of dispatching module. Changing to a structure which permits
procedures, since it contains all of the algorithms used for autonomy removes this problem while also eliminating the
control of the vehicle and the sensors. Its knowledge sources architecture sensitivity of the design.
and blackboards are strongly focused on detection and The concept of transmitting event data to more than one
processing of information affecting the routine execution of knowledge source makes the design easily adaptable to other
commands posted by the middle level. It contains the architectures, also. It can be installed in a system in which all
characteristics identifying detected objects as belonging to one knowledge sources are located in separate computers and
of the specified classes, and the appropriate reaction. In the communicate via a bus, or where all knowledge sources are in
case of obstructions, for example, the classes distinguished are the same computer and each is registered to be activated when
edge, wall, slope, or object. The reactions distinguished are specific kinds of events occur, or any combination of the two.
small maneuver, large maneuver, or pause. Corresponding In the case where all knowledge sources reside in the same
sensor activities are provided: narrow search or wide search processor, the operation will take the form that the blackboard
(vertical, horizontal, or both). is physically realized in a buffer, and each knowledge source
This design is unique in its combination of techniques. It is dispatched with a message about the buffer. This way, the
uses the blackboard approach, extended to encompass a more buffer may be distributed, with different memories for
complete form of the object-oriented paradigm, and a parti- different types of events. With this change, the system has
tioning of system functions by response time, to permit been made completely architecture-independent. Arbitration,
efficient operation. As a result, the system has a unique in this case, is limited to access to the transmission medium or
structure, being multidimensional. Since the design is inde- local resource, and to the adoption of the implications posted
pendent of hardware configuration, languages, number and by a given knowledge source. This is done by comparison to
types of knowledge sources, and knowledge representation the rules for strategy and tactics, and uses the implications
within a knowledge source, it is easily extended and easily derived by all knowledge sources. All knowledge has been

Authorized licensed use limited to: Bahauddin Zakariya University. Downloaded on March 11,2023 at 08:26:35 UTC from IEEE Xplore. Restrictions apply.
 ~

36 IEEE JOURNAL OF OCEANIC ENGINEERING, VOL. 13, NO. 2, APRIL 1988

Mlsdon Abstracl
I
(starto, finish0, surveyl:O, survey2:0, v i s i t 0 )
I I I
Design

-
I
lalternative 1 cf-’
,/
Design Suqvey 1 /

I (sensor)
LExternal
? \ (avoid8nce manewer) \
I
I Ewnt\ \
Internal
Event ,
f 5
\

Finish a t :
x1, Y 1 x2, Y 2
Actual: T+

Prot-Segment List 1
+-I---
10
7 I
I 8 9

I ~~
Execute Command Motion Sequence r
Conrrete
tarlierd Later

Fig 22 Problem solving by successive refinement.

compartmentalized, including both the expertise and the starts with a description of the mission and an ordering done at
operational knowledge, so that object-oriented design tools the supervisor level and transferred to the medium-term level.
such as Smalltalk, Flavors, and LOOPS can be used. Further, The method of illustration adopted is to show how detailed
this allows a newer operating system to be used, wherein planning for the mission shown in Fig. 20 is effected and
programs in any language can be called from the language of modified. The segment names used therefore reflect areas of
the knowledge-based system, so that the overall system is activity shown in that figure.
made potentially language-independent. Both figures reflect the same basic sequence of transitions in
2) Least Commitment and Alternatives: The concept of the planning cycle: mission-design-procedure-segment. The
multiple blackboards is supplemented by two other features first of these is the current projection of the sequence of
which make this design unique in real-time systems. These are execution of discrete packets of activity which constitute the
the use of a least commitment strategy and the existence of a components of the mission.
reasonable set of alternatives at every level and at all times. These data are stored and processed at the mission level,
The power of this approach is especially clear in the operation where all of the tactics information is stored. The data include
of the middle level, which handles analysis and execution for maneuvers for known features of the terrain. In effect, these
events and processes for which the required response time is data reflect both known facts and assumptions. The assump-
on the order of seconds. This level is labeled medium-term tions constitute a set of facts subject to retraction, thus giving
response in Fig. 19. The structure and operation of the system rise to the truth-maintenance problem typical of expert system
at this level will be used as an illustration of these techniques. operations.
Two approaches to planning and execution are shown in The second level, the design level, is the level at which
Figs. 22 and 23. These are the methods stressed in Hayes- planning of the sequence of steps to be used for each packet of
Roth. The first is called the successive refinement method. activity (such as “move to x,” or “survey y ” ) is accom-
The second is called temporaUspatia1. In either case, execution plished. The survey packets make good illustrations, because

Authorized licensed use limited to: Bahauddin Zakariya University. Downloaded on March 11,2023 at 08:26:35 UTC from IEEE Xplore. Restrictions apply.
 SIMMONS AND CHAPPELL: ARTIFICIAL INTELLIGENCE 37

Mission Abstrat

(start(), finish(), survey l:O, survey20, visit())

I I
3
Design
/’

alternative 1

Design Survey 1

Sensor: 9

Prc

Start at:
x1, Y1
Plan : T sec Actual : T+ b

10
2

* * *
Segment Possible Change 5I

I 1

Execute Command Motion Sequence

Concrete
Earlier + m- Later

Fig. 23. Problem solving-spatial/temporal ordering.

they consist of specific combinations of linked segments which however, is the inclusion of specificially stated alternatives
cover the desired area with a resolution dependent on the which are active at all times. This is represented as a third
amount of detail desired. dimension, directed into the page of the drawing. It transforms
The procedure level generates the sequence of motion the blackboard design to a design which can encompass three
commands required. Basically, this is the same as the control or more dimensions. At the “design” level, the alternatives
language of the “turtle” used in Lisp and Logo to introduce include such things as
children to the concept of programming: move xl, turn y deg., the ability to change the design of a survey pattern to
move x 2 , etc., until it is supplemented by the added
avoid an obstacle (or to examine in detail an object of
sophistication needed for maneuvers, as required by the
interest);
environment and the desired physics of the vehicle motion
the ability to adapt the design to variation, such as adding
(speed, turn rate, etc.).
a task, is needed to comply with a change order, such as
The “lowest” level of activity, the segment level, repre- adding a visit to a cluster, caused by the “notice cluster”
sents execution in “logical” time segments, where a logical event shown in Fig. 23.
time segment is defined as the time duration of a steady-state
of activity (forward motion, for example). The “notice cluster” event is the kind of activity used by
Thus far, the design is two-dimensional, as are other current Hayes-Roth to demonstrate the utility of the multiple-black-
blackboard systems. The basic dimensions are board control architecture.
Alternatives are also stored at the procedure level, as
vertical: level of abstraction: concrete at the bottom, and
indicated. The alternatives defined at this area, however,
abstract at the top;
reflect mostly the kind of changes precipitated by the
horizontal: time of execution: current time at the right,
occurrence of external events. Typically, they would consist
later time at the left.
of revising the instructions for the current segment to provide
The most significant new feature of this part of the design, for the incorporation of an avoidance maneuver. However,

Authorized licensed use limited to: Bahauddin Zakariya University. Downloaded on March 11,2023 at 08:26:35 UTC from IEEE Xplore. Restrictions apply.
 38 IEEE JOURNAL OF OCEANIC ENGINEERING, VOL. 13, NO. 2, APRIL 1988

activity may also be triggered by a command to interrupt or card stack program on the Macintosh, because the “alterna-
change a path because of a modified pattern that the mission or tives” dimension can be thought of as a set of cards which can
design level has created. be selected in any order. Therefore, execution of the mission
The segment path, when executed, does not create any involves creating a thread through the available sets of cards
design modifications itself. It is solely concerned with such that the path may connect to any sequence needed. As a
recording or forwarding the specific control signals needed in result, the initial estimates of distance, time, and energy are
the current execution frame. However, it has available to it changed, and the actual completion time is changed to the
reflex responses, and these can cause switching to an estimated time, plus or minus delta, as indicated.
alternative plan, as indicated by the alternate line of execution Minimizing the commitment to detailed plans, combined
shown in the background. At this level, the system incorpo- with making available alternatives at every level, also mini-
rates “canned” responses to various kinds of object or mizes the problem of truth maintenance. In the event that
obstacle which might be encountered. Initially, this consists of certain assumptions must be changed as a result of encounters
the shape classifiers “edge” for a narrow object, “block” for in the environment, retracing the steps already taken is
an object of larger extent (but with visible edges), “wall” for simplified. Parts of the path may be retraced and redone
an extended object, and “slope” for a smooth transition in because dependencies are stored, and the alternatives are
shape. Each of these objects carries with it a definition of the known. This operation is the same as the automatic truth
characteristics of the sensor information which indicate its maintenance systems of KEE, ART, and Nexpert. In general,
presence, and a set of avoidance commands. In the case of the however, there will be a limit on the amount of retraction and
“wall” object, there is an alternative which consists of revision allowed, using the heuristic of “point of no return,”
pausing to collect more data (perhaps including retreat and/or assuming that any segment successfully completed need not be
initiating a search for a top or an edge). retraced, so that the data for it can be discarded. This also
The sources of the events which trigger use of the minimizes the processing required to modify the set of facts
alternatives are sensors for external events, and processing of describing the environment, and hence the processing associ-
sensor and state data, for external events. In both cases, the ated with truth maintenance.
arrows indicate that the occurrence of the event causes activity The numbers shown in Figs. 22 and 23 indicate the order in
to propagate through all levels. As indicated by the discussion which the tasks would be executed, as a function of the
of the avoidance maneuver, this activity will encompass scheduling algorithm used. The numbers in Fig. 22 are the
immediate action while replanning is taking place. The order resulting from the use of the successive refinement
mechanism for recording all of this activity is the set of algorithm described in Hayes-Roth, and the numbers in Fig.
blackboards or communication media linking knowledge 23 result from use of the temporaUspatia1 algorithm. The
sources at all levels, as well as all knowledge sources at a successive refinement algorithm favors the results generated at
given level. The interaction of levels of abstraction is indicated the highest level of abstraction, with the least variability to
by the fact that the “notice cluster” action occurs at the level, account for the real world. The spatial/temporal algorithm
but the design activity is affected by it, via a notice procedure. favors the segment level, where the execution is actually
This linkage illustrates the value of the least commitment occurring, and where the impact of the real world is significant
strategy. In the event that the new activity should logically and immediate. The system uses both algorithms, in the roles
proceed before the next survey pattern starts, the design level appropriate to them, but with the possibility to switch between
simply notifies the procedure level of the postponement, and them. That is, the successive refinement algorithm is the basic
inserts the new activity ahead of it. Since no command strings method used at the supervisor level, and the spatial/temporal
have been loaded, the impact of the change is minimal. algorithm is best used at the lower levels. In this approach,
This mode of operation allows the system to act in an task designs are not necessarily completed or executed in the
opportunistic way, without the inefficiency of having to order used by the initial program created by the supervisor
discard instructions already generated. The initial mission module, but in the order determined by the temporal/spatial
specification, divided into the logical groups on which algorithm. Further, segment design is not finalized until the
segments are based, constitutes a wish list, and an ordering preceding task has been initiated, so that the actual procedure
which ensures completion of the mission with a reasonable can reflect variations caused by intervening events. This
reserve for fuel and pattern power. It includes a set of allows the least possible commitment to be made prior to
additional tasks which may be needed (for example, if specific actual need. The system will continue to use a mixture of the
actions are desired on encountering specific objects) and algorithms, to allow for alternation between them whenever
enough of a model of the environment to cover typical possible.
obstacles or terrain features which may require immediate The heuristic which is used to cause switching between the
response, with the appropriate response also specified. Fortui- algorithms is based on the time available to complete the plan.
tous events, such as the “notice cluster” event, can be The allowable planning time is a part of the mission
included, so that the tasks can be changed to provide more specification, which allows the planning process to be made
value, if the opportunity arises. The list also includes a list of context-sensitive, and to interact with response-time require-
things “not doable,” as well as the “doable” activities in the ments. It allows the system to adjust both the path and the
task descriptions. planning activity to account for the nature of the problem and
The overall operation resembles the operation of a Hyper- the capability of the vehicle. If the planning time is greater

Authorized licensed use limited to: Bahauddin Zakariya University. Downloaded on March 11,2023 at 08:26:35 UTC from IEEE Xplore. Restrictions apply.
 SIMMONS AND CHAPPELL: ARTIFICIAL INTELLIGENCE 39

than a specified limit, then the successive refinement strategy

is used, to obtain a more nearly optimum solution. If it is less
than the limit, then the spatial/temporal algorithm is used. The
difference is that, in the spatiaUtempora1 algorithm, knowl-
edge sources are dispatched in the order in which they request
control, so there is no guarantee that the final plan will in any
way be optimum.
The results of an execution using each of these algorithms,
in steps of operation of the inference engine, are shown in
Figs. 24 and 25. These figures show the activity occurring on
the execution blackboard. The gaps shown result from activity Seqment
4L
7‘
on the control blackboard, which is not illustrated. Fig. 24 1
shows results using the successive refinement solution, and 10 20 30 40 50 60 70
Fig. 25 shows results using the spatial/temporal (or “LIFO”) Fig. 24. Successive refinement solution
solution. The execution to which the figures apply is the
mission shown in Fig. 20.
Using the successive refinements technique, there is a
general stepping transition from the mission level down to the
segment level, with very little stepping back and forth between
levels. The mission to procedure to design transitions shown in

K\b
23 45
Desig~r
the “5-6-8” sequence reflect the ability of blackboard control
architecture to blend both top-down and bottom-up strategies,
as well as the ability of blackboard control architecture to Procedure 5 in
temporarily interrupt the successive refinement strategy.
Using the spatial/temporal technique, the graph has a Segment
’ 8 22
marked sawtooth shape, reflecting the lack of a control
I
heuristic for focusing the attention of the system at a given I
IO 2U 30 40 50 60 70
level of abstraction until a specified sequence is complete.
Comparing the sequences reveals interesting possibilities. Fig. 25. SpatiaUtemporal solution.
The spatial/temporal strategy produces the first segment,
therefore the first motion command, at cycle 18, while the because it allows for a mixture of technologies, and can easily
successive refinement strategy does not produce a command improve. It is domain-independent, implementation-indepen-
until cycle 49. This shows that the successive refinement dent, and architecture-independent, a notable combination of
strategy is not applicable to situations in which the reaction features. A combination of techniques and tools of this kind is
time is short, as in the case of encountering an obstacle. This found in all of the more successful systems implemented to
leads to the concept of having both strategies available: the date.
spatiaUtempora1 for replanning “locally, ” in the current
segment, and the successive refinement for replanning the IX. RESEARCH ISSUES
remainder of the mission, beyond the current segment. This The major research issues existing in AI today concern the
approach also exploits the best features of both approaches. internal structure of the machine used to implement the
The successive refinement strategy encompasses knowledge of knowledge-based system (e.g. von Neumann versus neural
the remainder of the mission as a whole (the “big picture”) net) and the relationship between the system architecture and
and therefore can provide judgment on viability. The result is the structure of the problem (e.g. single processor versus
to minimize wastefulness in “backtracking in motion” by multiple processor, or centralized versus blackboard).
maximizing the use of “backtracking in thinking,” prior to the The kind of machine used has become an issue because of
issuance of motion commands, as described by Chappell. difficulties in getting the standard machine to effectively
Since the spatiaMempora1 technique is subject to the genera- mimic important aspects of human behavior. One very
tion of wasteful motion, the mix is appropriate. The heuristics important example of a difficult behavior to emulate is the
which cause the switching to occur are supplied externally, human ability to work with partial information. None of the
prior to mission execution. They are constraints which include current approaches, using the von Neumann architecture,
priorities for the separate parts of the mission, preferred provide an adequate solution to this problem. However, an
choices, and rules for the manner in which segments are impressive success has been achieved using neural net
modified. machines, an emergent technology. It appears likely that the
3) Application Summary: All of these features combine to neural net machine will replace both current hardware and
make EAVE a notable blend of the methodologies of AI to current software for realization of expert systems, recognition
achieve a design which reflects human behavior in a sensible systems, signal detection and classification systems, and all
way, but without the pitfalls often associated with attempts to other systems which involve emulation of the human abilities
replicate the behavior of an “expert.” It is also noteworthy to detect, identify, and classify, using partial information in a

Authorized licensed use limited to: Bahauddin Zakariya University. Downloaded on March 11,2023 at 08:26:35 UTC from IEEE Xplore. Restrictions apply.
 40 IEEE JOURNAL OF OCEANIC ENGINEERING, VOL. 13, NO. 2, APRIL 1988

noisy environment. Software is affected because “program- more widely available, using traditional machines and special-
ming” in the current sense is meaningless for neural net ized Lisp machines. The mixture of techniques now provided
machines-they learn directly from the expert, typically using in commercial tools allows a more accurate modeling of the
the form of teaching called “supervised learning” and using nature, the organization, and the operation of human intelli-
an algorithm to control internal organization. The major gence. This advance has been the result of very large advances
problem with these systems at present is the lack of tools to in hardware technology, which have drastically reduced the
allow them to be specified accurately. cost of components, especially memory, while simultaneously
The issue of system architecture as compared to the increasing processing speed. The major problems in using the
structure of the problem being solved has become an issue tools effectively have been a) selection of a suitable problem
because of and b) elicitation of the relevant knowledge. No simple
solution exists for the problem of selecting a problem. The
the slowness of execution of programs on standard,
best that can be done is summarized in the guidelines given in
centralized machines;
Section V. As a result of the elicitation problem, it often
the independent nature of the various kinds of knowledge
happens that personnel trained in anthropology, linguistics,
and the difference in processing appropriately for the
and cognitive psychology are more effective at knowledge
various kinds.
engineering than personnel trained in computer science. In
It is clear that, for rule-based systems, ‘‘real-time operation,” general, when implementing a system, the more sophisticated
in the sense of military sensors, for example, is not now tools will be more useful, but some ability to do original
possible in a single processor. It is also clear that many parts of programming will be required. To allow for this and also to
the knowledge-based system could be executing their opera- allow for unexpected difficulty in reproducing the reasoning
tions simultaneously, in a truly parallel process, because they process, the problem selected should be as simple as possible.
represent different kinds of knowledge. For example, the rules At the same time, investment in hardware and software should
applicable to predicting weather are independent of selection be minimized as much as possible, so that the rapidly
of the type of resources required to prosecute a mission advancing technology can be exploited to the best advantage.
successfully. Therefore, there is much interest in architectures This is especially true in view of the new types of machines
which supply parallel processing, and in a design algorithm now appearing in the marketplace. Past successes and emer-
which effectively exploits this capability. Candidate system gent technology combined promise the possibility of a totally
architectures include networks of general-purpose machines, new generation of systems which will make the machine quite
and networks of machines which specialize in performing the capable of accurately emulating complex human behavior while
same operation on large quantities of data. The first kind of being built of very small and inexpensive components. The
computer is referred to as the distributed computer. The payoff is becoming sufficiently attractive to encourage use of
second kind is referred to as the massively parallel computer, a the technology for all kinds of tasks which involve heuristics
label that is sometimes applied to neural net machines, and human judgment. The practicality of a “knowledge
although the neural net machine is more commonly referred to amplifier” is now real, and its availability promises to create a
as the connectionist machine. The major problem on all powerful extension of human capability and performance.
machines other than the neural net machines has been
development of an operating system. The operating system for REFERENCES
a distributed system of any kind must be especially concerned M. Minsky, “A framework for representing knowledge,” MIT for
ONR, June 1984 (AD-A01 1168).
with a “reasonable” allocation of multiple resources. Design H. J. Berliner, “A prepresentation and some mechanisms for a problem
of a suitable resource allocation algorithm has been a major solving chess program,” CMU, May 1975 (AD-A012-917/1).
problem and has been complicated by the inclusion of H. J. Berliner, “BKG, A program that plays backgammon,” CMU,
July 1977 (AD-A043453/0).
somewhat arbitrary constraints (such as requiring memory to H. J. Berliner, “The B* tree search algorithm, A best first proof
be shared, and not allowing programs to be passed with data). procedure,” (interim report), CMU, Apr. 1978 (AD-A059-391/3).
One simple solution to this problem is to exploit the functional “Meta-rules: Reasoning about controls,” AL Lab, MIT 56789 (AD-
A084639).
independence of the resident modules. This is particularly “Constructing superview,” Dept. of Decision Scienses, Wharton
attractive because it allows the advantages of the blackboard School, U. of Penn. (AD-A093878).
type of system to be exploited. In the blackboard implementa- “Natural language communication with machines: An ongoing goal,”
BBN Rep. No. 5375, July 1983 (an excellent example of the goals
tion, however, there is no assurance that each processor will described in this report is given by the achievements of Englemann in
be used to its full capacity at all times. The assurance that the the TARPS system developed at MITRE Corp.) (AD-A131320).
system can provide needed solutions within strict time J. J. Richardson, “Artificial intelligence: An analysis of potential
applications to training, performance measurement, and job perform-
constraints may, however, override these objections. In ance aiding,” prepared for the Air Force Human Resources Labora-
addition, the concept of a network of specialists invites the use tory, Air Force Systems Command, Brooks AFB, TX., Sept. 1983
of processors best fitted to specific parts of the problem, such (AD-A133592).
Alphatech, Inc., Status report on contract to model the decision-making
as neural net machines for classification. process of submarine warfare commanders, Feb. 1983 (AD-A133476).
X. SUMMARY C. Hewitt and P. DeJong, “Analyzing the roles of descriptions and
actions in opens systems,’’ MIT AI Lab Rep. AIM 727, Apr. 1983
The field of AI has produced tools and methods which have (essentially a proposal to recreate Smalltalk-80, with object behavior
been effective in capturing human expertise and making it descriptors) (AD-A 133614).

Authorized licensed use limited to: Bahauddin Zakariya University. Downloaded on March 11,2023 at 08:26:35 UTC from IEEE Xplore. Restrictions apply.
 SIMMONS AND CHAPPELL: ARTIFICIAL INTELLIGENCE 41

[ I I] R. A. Dillard, “Representation of tactical knowledge shared by expert [41] S. G. Chappell, “A blackboard system for context sensitive mission
systems,” NOSC TD 632, Oct. 1983 (AD-A136875). planning in an autonomous vehicle,” presented at 5th Int. Symp.
[ 121 Alphatech, Inc., “Cognitive simulation of anti-submarine warfare Unmanned, Untetethered Submersible Technol., June 22-24, 1987.
commander’s tactical decision process,” NOSC TR 191, Aug. 1983 [42] E. Charniak and D. McDermott, Introduction to Artificial Intelli-
(AD-A138849). gence. Reading, MA: Addison-Wesley, 1985.
[I31 C . Hewitt and H. Lieberman, “Design issue in parallel architectures [43] W. F. Clocksin and C . S . Mellish, Programming in Prolog. New
for artificial intelligence,” prepared for ARPA, Nov. 1983 (AD- York: Springer-Verlag, 1981.
A 14282). 1441 R. Davis and D. B. Lenat, Knowledge-Based Systems in Artificial
1141 D. R. Brown, R. H. Monahan, and W. T. Park, “Artificial Intelligence. New York: McGraw-Hill, 1982.
intelligence/robotics applications to navy aircraft maintenance,” pre- 1451 S. R. Delaney, Babel-17. New York: Ace Books, 1966.
pared for the David W. Taylor Naval Ship Research and Development [46] A. Elithorn and R. Banerji, Artificial and Human Intelligence.
Center, Apr. 1983 (AD-A143219). North Lolland, 1984 (on order).
[ 151 R. F. Stengel, “Artificial theory and reconfigurable control systems,’’ [47] P. A. Eschholz, A. Rossa, and V. P. Clark, Language Awareness.
June 1984 (AD-A143766). New York: St. Martin’s Press, 1974 (no deep theory, just a set of
[ 161 “Some scientific subroutines in LISP,” AL Lab, MIT. (AD-A147889). interesting examples, and fun to read).
[ 171 “The documentation assistant: An intelligent system for documenta- [48] E. A. Esper, Analogy and Association in Linguistics and Psychol-
tion,” AI&DS, Mountain View, CA (AD-A154709). ogy. Athens, GA: University of Georgia Press, 1973.
[ I 81 “The design and implementation of an object-oriented, production rule [49] E. A. Feigenbaum and P. McCorduck, The Fifth Generation.
interpreter,” Naval Postgraduate Schools, 1984 (AD-A154769). Reading, MA: Addison-Wesley, 1983.
[I91 “GOAL: A goal oriented command language for interactive proof [50] E. A. Feigenbaum and A. Bar, The Handbook of Artificial
construction,” Standford AI Lab (not for the faint-hearted; concerned Intelligence, vol. I. Stanford, CA: Heuristic Press. Copyright by
with formal (symbolic) logic and the use of command household tools Kaufman Press, Los Altos, CA; published by William Kaufmann, Inc.,
such as the Takeuchi function and Ramsey’s theorem; interesting for 1982.
sophisticated students of math) (AD-A155093). [51] E. A. Feigenbaum, and P. R. Cohen, The Handbook of Artificial
1201 NOSC; “Experimental modeling of tactical data systems,” TD 548, Intelligence, vol. 111. Stanford, CA: Heuristic Press. Copyright by
Oct. 1982 (AD-B070623). Kaufmann Press, Los Altos, CA; published by William Kaufmann,
[21] S. M. McFadden, and M. M. Taylor, “Human limitations in towed Inc., 1982.
array sonar recognition,” prepared for the Defense and Civil Institute [52] J. A. Fishman, Sociolinguistics, A Brief Introduction. Rowley,
of Environmental Medicine, Ontario, Canada, May 1984 (the first MA: Newbury House, 1972.
attempt that I have seen to describe the nature of the problem of [53] J. A. Fodor, The Language of Thought. New York: Thomas Y.
significant features sensed by humans, so that signal processing Crowell, 1975 (an attempt to describe language in the “procedure”
operations could be optimized to provide maximum assistance to the terms of a computer, so that it can be created by a machine).
human) (AD-B084683). [54] A. Goldberg and D. Robson, Smalltalk-80, The Language and Its
[22] “A process for the rapid development of systems in support of manager Implementation. Reading, MA: Addison-Wesley, 1983.
decision-making,’’ AI Lab, MIT (DN83-0045). [55] A. Goldberg, Smalltalk-80, The Interactive Programming Environ-
[23] “Artificial intelligence and robotics,” M. Brady, AI Lab, MIT (DN84- ment. Reading, MA: Addison-Wesley, 1983.
0500). [56] J. H. Greenberg, Anthropological Linguistics: An Introduction.
[24] “Knowledge-based expert systems for off-shore engineering,” AI Lab, New York: Random House, 1968.
MIT (DN84-0051). 1571 R. K. Guha, “Design of a user microprogramming supporter system,’’
[25] R. Davis, “Diagnosis via casual reasoning: Paths of interaction and the in Proc. 15th IEEE Computer Soc. In/. Conf., 1977.
locality principle,” AL Lab, MIT (SN84-0017). [58] R. A. Hall, Jr., Soundand Spelling in English. Philadelphia: Center
[26] G. Agha, Actors. Cambridge, MA: MIT Press, 1986. for Curriculum Development, 1961.
[27] I. Asimov, P. S. Warrick, and M. H. Greenberg, Eds., Machines That [59] J . Haugeland, Artificial Intelligence, The Very Idea. Cambridge,
Think, Stories about Robots and Computers. New York: Holt, MA: MIT Press, 1985.
Reinhardt and Winston, 1983. [60] R. Hayes-Roth, D. A. Waterman, and D. B. Lenat, Building Expert
[28] C. Ballard and C. M. Brown, Computer Vision. Englewood Cliffs, Systems. Reading, MA: Addison-Wesley, 1983.
NJ: Prentice-Hall, 1982. [61] B. Hayes-Roth, “A blackboard architecture for control,’’ Artifcia1
[29] H. J. Berliner, “Experiences gained building and testing a chess Intelligence, vol. 26, pp. 251-321, 1985.
program,” in Proc. IEEE, Syst. Sri. Cybernetics Conf. (Pittsburgh, [62] P. Henderson, Functional Programming-Application and Imple-
PA), Oct. 14-16, 1970. mentation (Prentice-Hall International Series in Computer Science).
[30] H. J. Berliner, “The use of domain-dependent descriptions in tree- Englewood Cliffs, NJ: Prentice-Hall, 1980.
searching,” in Perspectives in Computer Science. London: Aca- 1631 C. F. Hockett, Language, Mathematics and Linguistics. The
demic Press, 39-062, 1977. Hague, Netherlands: Mouton & Co., 1962.
[31] H. J. Berliner, “Computer games: The computations of judgement,” [64] R. C . Holt, E. D. Lazowska, G . S. Graham, and M. A. Scott,
Computers and People (USA), vol. 30, nos. 3-4, Mar.-Apr. 1981. Structured Concurrent Programming with Operating Systems Ap-
[32] H. J. Berliner, “Backgammon computer program beats world cham- plications. Reading, MA: Addison-Wesley, 1978.
pion,” Artificial Intelligence (Netherlands), vol. 14, no. 2 , Sept. [65] G. Krasner, Smalltalk-80, Bits of History, Words of Advice.
1980. Reading, MA: Addison-Wesley, 1983.
[33] H. J. Berliner, “Computer backgammon,” Sri. Amer., vol. 242, June [66] C. Levi-Strauss, The Savage Mind. Chicago: Univ. of Chicago
1980. Press, 1962 (original French edition), 1966 (English edition).
[34] H. J. Berliner, “Chess,” doctoral thesis, CMU, Mar. 1974. [67] R. E. Levitt, “Using knowledge of construction and project manage-
[35] A. Borning, “The programming language aspects of Thinglab, A ment for automated schedule updating,” Stanford and Intellicorp,
constraint-oriented simulation laboratory,” A C M Trans. Program- 1985.
ming Languages Syst., vol. 3, no. 4, Oct. 1981. [68] R. P. Lippmann, “An introduction to computing with neural nets,”
[36] A. Borning, “Thinglab-A Constraint-oriented simulation labora- IEEE ASSP Magazine, vol. 4, no. 2, Apr. 1987.
tory,” Ph.D. dissertation, Stanford University, in 1979; available from [69] J. McCarthy, “Recursive functions of symbolic expressions and their
University Microfilms International, Ann Arbor, MI. computation by machine,” Commun. Ass. Comput. Mach., vol. 7,
[37] R. J. Brackhaman and H. J. Levesque, Eds., Readings. Los Altos, pp. 184-195, Apr. 1960.
CA: Morgain Kaufmann Publishers, 1985 (not for the novice, but it [70] W. S . McCullock and W. Pitts, “A logical calculus of the ideas
gets more readable after the first two papers). emmanent in neural nets,” Bull. Math. Biophys., vol. 5, 1943.
[38] E. C. Carterette and M. P. Friedman, Handbook of Perception, vol. [71] R. S. Michalski, J. G. Carbonell, and T. M. Mitchell, Machine
IV, Hearing. New York: Academic Press, 1978. Learning, A n Artificial Intelligence Approach. Palo Alto, CA:
[39] E. C. Carterette and M. P. Friedman, Handbook of Perception, vol. Tioga Publishing, 1983.
VII, Language and Speech. New York: Academic Press, 1976. [72] M. Minsky and S. Papert, Perceptrons, an Introduction to Compu-
[40] W. L. Chafe, Meaning and the Structure of Language. Chicago: tations Geometry. Cambridge, MA: MIT Press, 1969.
Univ. of Chicago Press, 1970. [73] M. Minsky, Ed., Semantic Information Processing. Cambridge,

Authorized licensed use limited to: Bahauddin Zakariya University. Downloaded on March 11,2023 at 08:26:35 UTC from IEEE Xplore. Restrictions apply.
 42 IEEE JOURNAL OF OCEANIC ENGINEERING, VOL. 13, NO. 2, APRIL 1988

MA: MIT Press, 1968 (see especially the thesis by Ross Quillian). Asa B. Simmons received the B.S degree in
M. Minsky, Society of Mind. Cambridge, MA: MIT Press, 1987. physics from the University of Texas (Austin) in
C. V. Negoita, Expert Systems, Fuzzy Systems. Menlo Park, CA: 1951 and the M.A. degree in linguistics from
Benjamin/Cummings, 1985. California State University (Fullerton) in 1975
H P. Nii, “Blackboard systems,” AIMagazine, vol. 7, nos. 3 and 4, He has had extensive experience in all aspects of
1986. system and software engineering, from mission
D A. Norman, Ed., Perspectives on Cognitive Science. Norwood, analysis to acceptance tests in the field. His work
NJ: Ablex Publishing, 1980. includes pioneering efforts in system definition and
D. A. Norman and D. E. Rumelhart, Explorations in Cognition. software design for phased-array radars and for
San Francisco: W. H. Freeman & Co., 1975. trainers. He also has experience in system manage-
J. Piaget and B. Inhelder, The Psychology of the Child. New York: ment, as head of a system design section, oversee-
Basic Books, 1969. ing both hardware and software design, and in the field as on-site supervisor
R. R. Restak, The Brain. Toronto, Canada: Bantam Books, 1984. for installation and test of air defense ground environment systems for the
E. Rich, Artificial Intelligence. New York: McGraw-Hill, 1983. NATO alliance. He also has extensive experience in system analysis,
R. C. Schank and P. D. Childers, The Cognitive Computer. modeling, and simulation, including event-driven models of systems such as
Reading, MA: Addison-Wesley, 1984. SUBACS. He has received commendations for excellence for software
J. R. Searle, Speech Acts-An Essay in the Philosophy of designs he has generated for various sonar trainers. At present, he is occupied
Language. Cambridge, England: Cambridge University Press, 1974. with a project to incorporate AI methods in an existing command and control
M. Stefik and G. Brbrow, “Object-oriented programming, Themes and system, with the introduction of neural net technology into company products,
variations,” AI Magazine, vol. 7 , nos. 3 and 4, 1986. and with courses in Lisp, AI, and neural nets. His interests include the nature
D. S. Touretzky, LISP, A Gentle Introduction of Symbolic of intelligence and the methodologies which may be useful in emulating it in
Computation. New York: Harper & Row, 1984. machines, where the machines and methods may be of conventional types, or
D. A. Waterman, A Guide to Expert Systems. Reading, MA: may be based on advanced computation models, as in the case of neural nets
Addison-Wesley, 1986. Mr. Simmons has presented papers in linguistics and has participated in
J. Weizenbaum, Computer Power and Human Reason. San Fran- workshops on AI. He also serves on the Corporate Committee for AI. He is a
cisco: W. H. Freeman & Co., 1976 (the inside story from the creator of member of the American Association for Artificial Intelligence, the Interna-
the Eliza program cited in the literature, with cogent comments on some
of the “reasoning” programs now available, such as the Macsyma
program of Symbolics).
W. E. Welmers, African Language Structures. Berkeley, CA:
*
tional Neural Net Society, and the Photographic Society of America.

Steven G. Chappell received the B S. degree in

Univ. of California Press, 1973. botany from the University of New Hampshire,
H. Whitaker and H. A. Whitaker, Eds., Studies in Neuro-Linguistics, Durham, in 1978. He received the M.S. degree in
vol. 1. New York: Academic Press, 1976. computer science from UNH in 1987. His thesis
T. Winograd, Language as a Cognitive Process. Reading, MA: was a prototype mission planner for an autonomous
Addison-Wesley, 1983. vehicle based on a dual blackboard architecture.
T. Winograd, Understanding Natural Language. New York: Aca- He is currently a Software Systems Engineer at
demic Press, 1972 (this is the world of “shrdlu” and the blocks, so the Marine Systems Engineering Laboratory
often cited in AI literature). (MSEL) at the UNH, which he joined in 1981
P. H. Winston, Artificial Intelligence, 2nd ed. Reading, MA: Since then, he has been designing, implementing,
Addison-Wesley, 1984. and maintaining software for MSEL’s experimental
P. H. Winston and K. A. Prendergast, Eds., The AI Business. autonomous vehicle (EAVE). His interests include autonomous systems,
Cambridge, MA: MIT Press, 1984. artificial intelligence programming techniques applied to real-world problems,
P. H. Winston and B. K. P. Horn, LISP. Reading, MA: Addison- real-time operating systems, parallel processing, and neural networks. He is
Wesley, 1984. currently workmg on a world model for the prototype mission planner and
systems software for the third-generation EAVE vehicle.

Authorized licensed use limited to: Bahauddin Zakariya University. Downloaded on March 11,2023 at 08:26:35 UTC from IEEE Xplore. Restrictions apply.