PUFF/VM PROJECT Section 6.4.5

-- Diagnose need for change in patient ventilation;

-- Identify indications for proceeding forward or back in the process of
weaning from a ventilator; and

-- Make interpretations in light of measured test results, patient history,
record of therapies and results of therapies and observations.

4. Implement rules for interpretation of respiratory insufficiency data with a
new heuristic interpretation system including the following major features:

-- Forms time-dependent hypotheses about the patient state;

~- Infers desired courses of action based on measured patient state,
observations, and expectations of future course.

-~- Uses models, both heuristic and mathematical, for generating an
expectation of the immediate patient course;

5. Create an advisory committee, including outside experts in clinical
medicine and computer science, to review the progress to date. They will
review conceptual formulations, system design, scope and detail of the
Clinical knowledge and system operation. The advisory group will be asked
to help to identify additional important considerations for the clinical
knowledge base and the computer implementation, suggest improved ways to
conceptualize or implement problems, and evaluate the soundness of the
results to date.

N. Significance

Science advances by quantitation and development of general theories. The
practice of medicine advances along one path by integrating quantitative
measurements and general theories into the routine of existing clinical practice.
The world of clinical medicine includes a complicated interaction among human
patients, complex physiology, and proud, human clinical staff. This project is
based on the assertion that good, quantitative measurements of physiological
state are useful if effectively related to the human and physiological
complexities of the clinical world. The vest possibility we see for making new
quantitative measurements far more generally useful in clinical medicine lies in
knowledge-based interpretation of well understood physiologically relevant
measurements. The improved care of the sick patient is our objective. This
project, if successful, will directly improve the ability of the clinical starr
to properly diagnose and manage the patient with respiratory insufficiency. It
will lay the foundation for extension of successful methodologies of
interpretation of the general problem of interpreting measurements of
physiological state.

Privileged Communication 201 J. Lederberg
Appendix I OVERVIEW OF ARTIFICIAL INTELLIGENCE RESEARCE

Appendix I

OVERVIEW OF ARTIFICIAL INTELLIGENCE RESEARCH

 

ARTIFICIAL INTELLIGENCE RESEARCH
What is it? What has it achieved? Where is it going?
Excerpt from a report by

Professor Edward A. Feigenbaum
Stanford University

May 1975

INTRODUCTION

In this briefing, these questions will be discussed as succinctly as
possible:

I. What is the scientific field of artificial intelligence research, as seen
fron various viewpoints? What are the general goals of the field?

If. What are its practical workins goals? What are some achievements relative
to these goals (cirea 1973)?

Til. What steps (new goals, problems, potential achievements) seem to lie ahead,
within a five year horizon?

ARTIFICIAL INTELLIGENCE (alias INTELLIGENT COMPUTER SYSTEMS):
General View;

Artificial Intelligence research is that part of Computer Science that is
concerned with the symbol-manipulation processes that produce intelligent action.
By “intelligent action" is meant an act or decision that is goal-oriented,
arrived at by an understandable chain of symbolic analysis and reasoning steps,
and is one in which knowledge of the world informs and guides the reasoning,

Some scientists view the performance of complex symbolic reasoning acts by
computer programs as the sine qua non for artificial intelligence programs, but
this is necessarily a limited view.

Yet. another view unifies AI research with the rest of Computer Science. It
is an oversimplified view, but worthy of consideration. The potential uses of
computers by people to accomplish tasks can be "one-dimensionalized" into a

J. Lederberg 202 Privileged Communication
OVERVIEW OF ARTIFICIAL INTELLIGENCE RESEARCH Appendix I

Spectrum representing the nature of instruction that must be given the computer
to do its job. Call it the WHAT-TO-HOW spectrum. At one extreme of the
Spectrum, the user supplies his intelligence to instruct the machine with
precision exactly HOW to do his job, step-by-step. Progress in Computer Science
Can be seen as steps away from that extreme "HOW" point on the spectrum: the
familiar panoply of assembly languages, subroutine libraries, compilers,
extensible languages, etc. At the other extreme of the spectrum is the user with
his real problem (WHAT he wishes the computer, aS his instrument, to do for him).
He aspires to communicate WHAT he wants done in a language that is comfortable to
nim (perhaps English); via communication modes that are convenient for him
(including perhaps, speech or pictures); with some generality, some abstractness,
perhaps some vagueness, imprecision, even error; without having to lay out in
detail all necessary subgoals for adequate performance ~ with reasonable
assurance that he is addressing an intelligent agent that is using knowledge of
his world to understand his intent, to fill in his vagueness, to make specific
his abstractions, to correct his errors, to discover appropriate subgoals, and
ultimately to translate WHAT ne really wants done into processing steps that
define HOW it shall be done by a real computer. The research activity aimed at
creating computer programs that act as "intelligent agents" near the WHAT end of
the WHAT-~TO-HOW spectrum can be viewed as the long-range goal of AI research,
Historically, AI research has always been the primary vehicle for progress toward

this end, though science as a whole is largely unaware of the role, the goals,
and the progress.

HISTORICAL TRACE
The working Goals of the science;
Progress toward those goals;

The root concepts of AI as a science are 1) the eonception of the digital
computer as a symbol-processing device (rather than as merely a number
calculator); 2) the conception that all intelligent activity can be precisely
described as symbol-manipulation. (The latter is the fundamental working
hypothesis of the AI field, but is controversial outside of the field.) The first
inference to be drawn therefrom is that the symbol-manipulations which constitute
intelligent activity can be modeled in the medium of the symbol-processing
capabilities of the digital computer.

This intellectual advanee--which gives realization in a pnysical system,
the digital computer, to the complex symbolic processes of intelligent action and
decision--with detailed case studies of how the realization can be accomplished,
and with bodies of methods and techniques for creating new demonstrations--ranks
aS one of the great intellectual achievements of Science, allowing us finally to
understand how a physical system can also embody mind. The fact that large
segments of the intellectual community do not yet understand that this advance
has been made does not change its truth or its fundamental nature.

Privileged Communication 203 J. Lederberg
Appendix I OVERVIEW Of ARTIFICIAL INTELLIGENCE RESEARCH

Three global "working goals" have dominated the AI field for the 17 years
of its existence, These are:

1. Understanding heuristic search as a processing scheme sufficient to account
for much intelligent problem solving behavior; and exploring the scope and
pervasiveness of heuristic problem solvinz.

2. Semantic information processing: developing precise formulations of
"understanding" by programs, and "meaning" of symbols that are input or
stored; the acquisition, storage, and deployment of knowledge of the world in
the service of symbolic problem solving.

3. Information Processing Psychology: developing precise models of human
behavior in synbolic-processing tasks.

The first two goals represent the fundamental paradigms that have dominated
the field. The third cuts across these orthogonally, and involves intense
interdisciplinary contact with Psychology, and Linguistics.

GOAL 1. HEURISTIC SEARCH, HEURISTIC PROGRAMMING, SYMBOLIC
PROBLEM SOLVING PROGRAMS

In the first decade, the dominant paradign of AI research was heuristic
searen, In this paradigm, problem solving is conceived as follows: A tree of
"tries" (aliases: subproblems, reductions, candidates, solution attempts,
alternatives-and-consequences, etc.) is sprouted (or sproutable) by a generator,
Solutions (variously defined) exist at particular (unknown) depths along
particular (unknown) paths. To find one is a "problem". For any task regarded
as nontrivial, the search space is very large. Rules and procedures called
heuristics are applied to direct search, to limit search, to constrain the
Sprouting of the tree, etc. While some of this tree-searching machinery is
entirely task-specific, other parts can be made quite general over the domain of
designs employing the heuristic search paradizn. Two notions are critical. The
first is that problem solvers generally face a "maze" of alternative courses of
decision and action that is huge compared with their processing resourees. The
second is the use of heuristic knowledge to steer carefully through large mazes
toward a solution seeking the plausible and potentially fruitful avenues,
avoiding the absurdities and the high-risk paths. Heuristic knowledge is usually
informal knowledge--to be distinguished from formal knowledge that is assertable
with the rigor of proof. Polya, the famous mathematician who wrote Patterns of
Plausible Inference and other books on problem solving, calls heuristic reasoning
"the art of good guessing." Heuristic knowledze is often "common sense" knowledge
of the world, rules-of-thumb for generally acceptable performance, or rules of
good practice in specific situations. When we speak of the "expertise" of an
expert, and the "good judgment" he brings to bear on complex problems in his
domain, we often are speaking of the heuristics he nas developed to search
effectively.

J. Lederberg 204 Privileged Communication
OVERVIEW OF ARTIFICIAL INTELLIGENCE RESEARCH ' Appendix I

Provocative essays by Polya notwithstanding, the first serious and detailed
studies of heuristic problem solving ever done by Science were done as AI
research in its first decade. As with any other science, progress came by the
detailed examination of specific cases, from which gradually emerged both a broad

picture of the nature of the phenomena being studied and, within this, more
formal theories for specific parts.

Three sub-goals of heuristic programming are discernable.

SUBGOAL 1A. Demonstrate sufficiency of heuristic search for
tasks of intellectual difficulty.

These heuristic programming efforts dealt with almost "pure" symbolic
reasoning tasks (i.e., tasks not requiring much coupling to real-world
knowledge), and used inference schemes that were either ad-hoc or of limited
scope. Notable successes during this "prove-the concept” phase were: the
Logic Theory Program, that proved theorems in Whitehead & Russel’s
propositional calculus; the Geometry Theorem Proving program, that proved
theorems in Euclidean geometry at a level of competence exceeding that of the
excellent high school geometry student; the Symbolic Integration program, that
Solved college freshman symbolic integration problems about as well as MIT
freshmen; chess-playing programs that play respectable "club player" C or B
Class chess; a checker playing program that was virtually unbeatable, except
by the country’s top few players (notable also for remarxable self-improvement
in performance by analysis of its own play and "book-move" good play); and a.
number of competent management science applications (assembly-—line balancing,
warehouse location, job-shop scheduling, etc.).

To recapitulate briefly: the key concepts are: search in problem
solving; and the use of generally informal knowledge to guide search
effectively. Tne AI community was the first to devote serious scientific
effort to developing the idea of the use of informal knowledge in problem
Solving, with notable successes. Few in Science recognize that this
achievement has been made and is ready for exploitation.

SUBGOAL 1B. Generality in Problem Solving Programs

Generality here means the use of a small set of problem solving methods
of wide applicability to solve problems of many different types. Each of the
problems posed is stated to the program in a particular representation (or
framework) witn which the set of methods is constructed to handle.

The subgoal of generality arises first as a reaction to the array of
"specialty" programs mentioned above; second, from the general observation
that the ability to do a wide range of tasks is a special touchstone of
intelligence; third, from a direct assessment that as tne diversity and
heterogeneity of the tasks handled by an agent increases, the likelihood that
it can do them all without intelligent action decreases; and fourth, from tne
argument that. any ultimate intelligent agent must have wide generality, since
it must take the world and its problems as they come without any intermediary,
maxing generality an important independent desideratum.

Privileged Communication 205 J. Lederberg
Appendix I . OVERVIEW OF ARTIFICIAL INTELLIGENCE RESEARCH

This subgoal was pursued with vigor for ten years in a number of
projects, was important for its feedback value in clarifying issues for the AI
field, and has temporarily (at least) been put back on the shelf as the field
begins to explore knowledge-based problem solvers and issues in the
representation of knowledge.

There were two discernable subthemes. The first was an attempt to
create abstract heuristic search methods that were divorced from any
particular content. Examples were: tne General Problem Solver, which used a
variant of heuristic search known as mean:-ends analysis; MULTIPLE, which
introduced adaptivity in the selection of what subproblem to choose "next" in
a search; and REF-ARF, which extended the generality of ordinary procedural

programming languages to include the embedding of non-procedural problems of
constraint satisfaction.

The second subtheme was the construction of theorem provers that take
problems expressed as theorems to be proved in the first-order predicate
calculus. This line of work was motivated by the (correct) observation that
the scope for representing real-world facts and situations in first-order
predicate calculus is very great; and by the invention of the resolution
method, a computational method for finding proofs for theorems in this
calculus. There has been continuous improvement on the basic method, taking
the form of proposing more powerful inference techniques, rather than the form
of specific ways for programs to adapt to particular poroblems. The very
strength of the formulation in terms of generality, namely its complete
homogenization of the particular task (all tasks are seen and dealt with in
the same logical formalism) turns effort away from how to exploit the
particularities of special classes of tasks. But it appears that only by
exploiting the particularities can significant reduction in search be
achieved. From a practical point of view the only proofs produced by such
problem solvers were "shallow" proofs.

Much of this line of research has been temporarily "shelved", awaiting
further knowledge on how best to represent knowledge for computer processing.
Problems that are essentially simple when represented in their "natural"
representation appear extraordinarily complicated when translated into first-
order predicate calculus. The current search for theorem provers using
higher-order logics is based not on the attempt to increase the raw expressive
power, so to speak, of first-order logic, but on the belief that naturalness
of expression will ultimately pay off.

SUBGOAL 1C: High-Performance Programs that perform at near-human
level in specialized areas

As the heuristic programming area matured to the point where the
practitioners felt comfortable with their tools, and adventuresome in their
use; as the need to explore the varieties of problems posed by the real-world
was more keenly felt; and as the concern with knowledge-driven programs (to be
discussed later) intensified, specific projects arose which aimed at and
achieved levels of problem solving performance that equalled, and in some
cases exceeded, the best human performance in the tasks being studied. The

J. Lederberg 206 Privileged Communication
OVERVIEW OF ARTIFICIAL INTELLIGENCE RESEARCH Appendix I

example of such a program most often cited in the Heuristic DENDRAL progran,
which solves the scientific induction problem of analyzing the mass spectrum
of an organic molecule to produce a hypothesis about the molecule’s total
Structure. This is a serious and difficult problem in a relatively new area
of analytical chemistry. The program’s performance has been generally very
competent and in "world’s champion" class for certain specialized families of
molecules. Similar levels of successful performance have been achieved by
some of the MATHLAB programs that assist scientists in doing symbolic
mathematics. The effectiveness of MATHLAB’s procedures for doing symbolic
integration in calculus is virtually unexcelled. Yet another example, with
great potential economic significance, involves a program for planning complex
organic chemical syntheses from substances available in chemical catalogs,

The program is currently being used as an "intelligent assistant" in a new and
complex organic synthesis.

GOAL 2. SEMANTIC INFORMATION PROCESSING (S.1I.P.)

The use of the term "semantic" above is intended to connote, in familiar
terns, something like: "What is the meaning of..." or "How is that to be
understood..." or "What knowledge about the world must be brought to bear to
solve the particular problem that has just come up? The research deals with the
problem of extracting the meaning of: utterances in English; spoken versions of
these; visual scenes; and other real-world symbolic and signal data. It aims
toward the computer understanding of these as evidenced by the computer’s

“Subsequent linguistic, decision-making, question-answering, or motor behavior.

Tous, for example, we will know that our "intelligent agent" understood the
meaning of the English command we spoke to it if: a) the command was in
itself ambiguous; b) but was not ambiguous in context; and c) the agent
performed under the appropriate interpretation and ignored the
interpretation that was irrelevant in context.

/In this goal of AI research, there are foci upon the encoding of knowledge
about the world in symbolic expressions so that this knowledge can be manipulated
by programs; and the retrieval of these symbolic expressions, as appropriate, in
response to demands of various tasks. S.I.P. has sometimes been called “applied
epistemology" or "knowledge engineering".

To summarize: the AI field has come increasingly to view as its main line
of endeavor: knowledge representation and use, and an exploration of
understanding (how symbols inside a computer, which are in themselves essentially
abstract and contentless, come to acquire a meaning).

To classify all of the current work into a relatively simple set of
Subgoals is a formidable and hazardous undertaking. Nevertheless, here is one
rough cut (stated for convenience as questions).

Privileged Communication 207 J. Lederberg
Appendix I . OVERYI2W7 OF ARTIFICIAL INTELLIGENCE RESEARCH

A. How is the knowledge acquired, that is needed for understanding and problem
solving; and how can it be most effectively used?

B. How is knowledge of the world to be represented symbolically in the memory of
a computer?

Bi. what symbolic data structures in memory make the retrieval of this
information in response to task demands easy?

C. How is knowledge to be put at the service of programs for understanding
English?

D. How is sensory knowledge, particularly visual and speech, to be acquired and
understood? How is knowledge to be applied to intelligent action of
effectors, such as arms, wheels, instrument controls, ete.

Significant advances on all of these fronts have been made in the last
decade. The area has a rather remarkable coherence--with individual projects
threading through a number of the goals stated above (this makes excellent
science and difficult exposition!)

GOAL 2A. Knowledge Acquisition and Deployment for Understanding
and Problen Solving

The paradigm for this goal is, very generally sketched, as follows:

a. a situation is to be described or understood; a signal input is to be
interpreted; or a decision in a problem-solution path is to be made.

Examples: A speech signal is received and the question is, "What was
said?" The TV camera system sends a quarter-~million bits to the
computer and the question is, "hat is out there on that table and in
what configuration?" The molecule structure-generator must choose a
chemical functional group for the “active center" of the molecular
structure it is trying to hypothesize, and the question is, "What does
the mass spectrum indicate is the “best guess“?"

b. Specialized collections of facts about the various particular task
domains, suitably represented in the computer memory (call these Experts)
can recognize situations, analyze situations, and make decisions or take
actions within the domain of their specialized knowledge.

Examples: In the CMU Hear-Say Speech Understanding System, currently
the Experts that contribute to the Current Best Hypothesis are an
Acoustic-Phonetic Expert, a Granmar Expert, and a Chess =xpert (since

J. Lederberg 203 Privileged Communication
OVERVIEW OF ARTIFICIAL INTELLIGENCE RESEARCH Appendix I

chess-playing is the semantic domain of discourse). In Heuristic
DENDRAL, the Experts are those that know about stability of organic
molecules in general, mass spectrometer fragmentation processes in
particular, nuclear magnetic resonance phenomena, ete.

For each of the sources of knowledge that can be delineated, schemes
must be created for bringing that knowledge to bear at some place in the on-—
going analysis or understanding process. The view is held that programs
Should take advantage of a wide range of knowledge, creating islands of
certainty as targets of opportunity arise, and using these as anchors for
further uncertainty reduction. It is an expectation that always some
different aspect provides the toe-hold for making headway--that is » that.
unless a rather large amount of knowledge is available and ready for
application, this paradigmatic scheme will not work at all.

Within this paradigm lie a number of important problems to which AT
research has addressed itself:

a. Since it is now widely recognized that detailed specific knowledge of task
domains is necessary for power in problen solving programs, how is this
knowledge to be imparted to, or acquired by, the vrezgrams?

al. By interaction between human expert and program, made ever more snooth
by careful design of interaction tecnniques, languages "tuned" to the
task domain, flexible internal representations. The considerable
effort invested by the AI community on interactive time-sharing and
interactive graphic display was aimed toward this end. So is the
current work on situation-action tableaus (production systems) for
flexibly transmitting from expert to machine details of a body of
knowledge.

a2. "“Custom-crafting" the knowledge in a field by the painstaking day-
after-day process of an AI scientist working together With an expert
in another field, eliciting from that expert the theories, facts,
rules, and heuristics applicable to reasoning in his field. ‘This was
the process by-which Heuristic DENDRAL’s "Expert" knowledge was built.
It is being successfully used in AI application programs to: diagnosis
of glaucoma eye disease, to treatment planning for infectious disease
using antibiotics, to protein structure determination using X-ray
erystallography, to organic chemical synthesis planning, to a military
application involving sonar signals, perhaps to other areas, and of
course to chess.

a3. By inductive inference done by programs to extract facts,
regularities, and good heuristics directly fron naturally-ocecurring
data. This is obviously the path to pursue if AI research is not to
spend all of its effort, well into the 21st Century, building
knowledge-bases in the various fields of human endeavor in the custon-
crafted manner referred to above. The most notable successes in this
area have been:

Privileged Communication 209 Jd. Lederberg
Appendix I OVERVIEW OF ARTIFICIAL INTELLIGENCE RESEARCH

a.

GOAL 2B.

..the Meta-DENDRAL program whicn, for example, has discovered the
mass spectrum fragmentation rules for aromatic acids from
observation of numerous spectra of these molecules--rules
previously not explicated by the DENDRAL chemists.

..a draw-poker playing program tnat inferred the heuristics of
good play in the game by induction (as well as by other modes,
including the aforementioned interaction with experts).

By processes of analogical reasoning, by which knowledge acquired
about one area can be used to solve problems in a another area if a
Suitable analogy can be drawn. Our human experience tells us that
this approach is rich in possibilities. One successful project can be
cited (and that is a limited success); a program that discovers an
analogy (in full-blown detail) between a theorem-to-be-proved in
modern algebra and another theorem in algebra whose proof is known.
The analogy is used to pinpoint from a large set of facts those few
that will indeed be relevant to proving the new theorem.

Representation of Knowledge

The problem of representation of knowledge for AI systems is this: if

the user has a fact about the world, or a problem to be stated, in what form
does this become represented symbolically in the computer for immediate or
later use? Three approaches are being pursued:

Bl.

the approach via formal logic. As mentioned before, first-order
predicate calculus was tried, but was found to be too cumbersome to
répresent ordinary situations and common-sense knowledge. Set theory and
higher-order logics are currently under examination as better candidates
to be a medium for homogeneous representation.

The ad-hoc approach. Most problem domains have a "natural”
representation that human experts use when operating in the domain.
Translate that representation fairly directly for the computer, and
tailor the information processes to work with it. This is the approach
commonly taken, in DENDRAL, MATHLAB, in chess playing programs, visual
scene analysis, and so on (almost everywhere). Though it gets the job
done, it creates serious problems for the cumulation of knowledge,
techniques, and programs in the seience because of the inhonogeneity that
arises therefroa throughout the collection of AI projects undertaken.

One way out of the dilemma is to do re search on the problem of
translation (by program) from one ad-hoc representation into another (the
so-called "shift of representation" problem). Little work has been done
on this problem, except one excellent "pencil-and-paper" exercise in

connection with a simple puzzle, and one subvrogram in DENDRAL (the
Planning Rule Generator, that translates mass spectral knowledge from its
form as fragmentation processes to a form useful for pattern matching) .

J. Lederberg 219 Privileged Communication
OVERVIEW OF ARTIFICIAL INTELLIGENCE RESEARCH Appendix I

B3. the approach via a "computable" semantic theory. In this approach,
computational linguists attempt to analyze the full range of actions,
actors, objects, and their relations, of which the common-sense world is
composed; then refine and formalize these into a useable computational
tneory for representing facts, utterances, problems, etc. The most
successful of these efforts is the Conceptual Parser (and its follow-on,
MARGIE, which successfully accomplishes English paraphrase and common-
sense inference).

In lieu of a tight, parsimonious computable semantic theory, other more
ad-hoc systems, known as semantic-net-memory models, have developed
experimenting with various sorts of actor-action-object-relation data
structures. Semantic-net-memory models have a ten year history relating
particularly to intelligent question-answering. Perhaps most successful of
these is the HAM program which combines ideas from semantic theory, semantic-
net-memory structures, and more traditional linguistic analysis (all in the
context of a rather good model of human sentence comprehension, validated with
dozens of careful laboratory experiments) .

GOAL 2C. Programs for Understanding English

One can readily observe that it will be almost impossible to disentangle
the skein of research on understanding natural language (English) from the
coordinate efforts in representation and deployment of knowledge. Most of the
state-of-the-art programs for understanding Englisn employ, in one form or
another, the basic S.I.P. paradigm outlined previously. These systems have
substantial linguistic components that are highly sophisticated compared with
anytning done in the past. All of them incorporate linguistic theory that has
an intimate and continuous tie-in between gramnar "Experts" and domain.
dependent "Experts", Although the domains about which they admit discourse
are still modest and discrete, they are many times richer than anything done
previously. The state-of-the-art is represented by the SHRDLU progran for
conducting a dialogue with a simulated robot about a world of blocks, boxes,
and pyramids on a table; and the Lunar Rocks program for conducting a dialogue
about properties of and transformations upon NASA moon-rock samples. The
SHRDLU program, for example, will carry out commands, answer questions, and
generally be aware of what it was doing, so as to answer "how" and why"
questions about its behavior.

The internal structure of these systems exhibits an interesting
evolution over the semantic-net-memory systems, and they appear to be a lone
way from the heuristic search schenes mentioned earlier. They are essentially
large programs written within a programming system that provides search and
matching capability. There is no factorization between a data base (i.e.,
semantic net) and a small set of methods that process the data base, Rather,
the entire system appears to be a large collection of special purpose prograns
for dealing with a multitude of special cases. They give the appearance of
being a highly distributed system, in which the intellizent action resides
throughout the entire program.

Privileged Communication 211 J. Lederberg
Appendix I OVERVIEW OF ARTIFICIAL INTELLIGENCE RESSARCH

GOAL 2D. Acquiring and Understanding sensory Data.

The goal here is to discover broadly applicable methods for extracting
from sensory data (chiefly visual and aural) the information that is
specifically responsive to users” needs. Two classes of needs may be noted:
the need to facilitate communication between man and machine; and the need to
apply computers to intrinsically perceptual tasks. The former is exemplified
by the desire to talk, rather than type, to computers; the latter is
illustrated by the task of automatically guiding an effector on the basis of
visual data. To satisfy either (or both) of these needs, it is necessary to

move from well-understood problems of sensing data to much more difficult
problems of interpretation.

SUBGOAL 2D1. Visual Scene Analysis. Computer-based analysis of visual
scenes has its roots in work on optical character recognition (early to mid-
Fifties) and by work in automatic photoreconnaissance. These tasks are
essentially two-dimensional. Little is lost by disregarding dimensions of
objects in a direction orthogonal to the picture plane.

AI research on scene analysis began in the early sixties with the work of
Roberts on pictures of polyhedra. This work (and its intellectual
descendants) differs from the earlier two-dimensional work in two major
respects: first, it explicitly considers, and capitalizes on, the three-
dimensional properties of objects and their perspective representations;
second, it utilizes a variety of special processing steps and decision-
making criteria, in contrast to the earlier template-match/classify
paradigm,

Robert’s work spawned five years of intensive research on pictures of
collections of polyhedra. One theme, centered on the archetypical question
"Is an edge present in a given (small) region of the picture?", led to the
development of edge detecting, contour following, and region finding
programs. A second theme, centered on teasing out the properties of
polyhedra and their representations, led to an elegant theory of permissible
representations of edges and vertices, and their relations to three
dimensional polyhedra - a theory not previously discovered by projective or
descriptive geometers.

Work in the polyhedral objects domain culminated in several programs capable
of describing, in more or less complete detail, pictures of complicated
collections of polyhedra, even taking into account shadows cast by these
objects. At the same time, more complicated types of scenes began to be
seriously studied. This has led to current interest in the use of color,
texture, and range data, and has stimulated interest in program
organizations capable of capitalizing on these multiple perceptual
modalities. For example, in one paradign perception is viewed as a problem-
solving process that uses many varieties of knowledge to select perceptual
operators, to guide their application to sensory data, and to evaluate the
results obtained therefrom.

J. Lederberg 212 Privileged Communication
OVERVIEW OF ARTIFICIAL INTELLIGENCE RESSARCH Appendix I

SUBGUAL 2D2. Speech Understanding. Research on computer recognition of
speech signal data began in the Fifties with work on the recognition of
isolated words. Some observations will be made here on the relation between
speech understanding research and the on-going body of AI research,

The fundamental idea driving research on speech understanding is that
"recognition" is impossible (in flowing natural speech) without
understanding, and that understanding is impossible without extensive
Knowledge about the domain of discourse. This view arises in part from the
observation that ambiguities and omissions at both the acoustic and semantic
level do not arise as bizarre or pathological exceptions but instead are
commonplace events. Speech understanding research thus relies heavily on
progress in the basic AI research problems of knowledze acquisition,
representation, and deployment. This situation in unlikely to change
regardless of advances in processing acoustic signals.

GOAL 2E. Intelligent control of Effectors

This goal concerns the creation of devices and control programs for
bringing about specified changes in the physical world. The effectors that
have attracted the most attention have been mechanical manipulators and mobile
vehicles but this has been largely a matter of experimental convenience. In
principle, they could as easily have been subsystems of spaceships or
manufacturing tools.

Early work in “intelligent” effectors dates back two decades, but
Systematic work did not begin until about 1965, at which time some progress
nad already been made in developing symbolic problem solving programs to
control effectors. Since then there has been considerable interest in
computer-controlled effectors because problems of effector control excite a
set of important issues for AI research. The following is a rough
characterization of the subgoals of work on effector control:

SUBGOAL E1. Monitoring Real-World Execution of Problem Solutions: The
special touchstone of effector control research is that a problem is never
"solved" until the real, physical world has been altered in a fashion that
Satisfies the task specification (in contrast to other problem solving
programs whose responsibility ends with the symbolic presentation of a good
solution). Tnus, an effector control program should ideally be prepared to
deal with any eventuality that affects the exeaution of a theoretically
correct solution, be it initial misinformation, accidental dynaaic effects,
ete. These demands strongly influence all levels of progran orzanization
and strategy. Problem solving and execution monitoring must be made to
interact intimately. The most advanced work of this type is probably the
STRIPS-PLANEX system (for the control of a mobile vehicle) that can detect
and sracefully recover from a wide variety of execution difficulties.

Privileged Communication 213 J. Lederberg
. So
Aposndix I OVERVIEW OF ARTIFICIAL INTELLIGENCE RESEARCH

J.

Lederberg 214

SUBGOAL E2. Modelling "Everyday" worlds: To control effectors by computer
requires that the computer have adequate models of everyday situations. It
has become important to model occlusion, obstruction, relative location,
etc., and tois has been done to the extent necessary to handle various
siuple manipulation and locomotion problems.

SUBGOAL E3. Planning in the face of uncertainty. Problem-solving programs
for the control of effectors that operate cn the physical world must be able
to work routinely with incomplete and inaccurate information. This creates
a need to do research on programs that can form contingency plans, can plan
to acquire information, can decide when to execute actions in the physical
world, even if the plan is incomplete, and so forth. Some research of this
type has been done.

SUBGOAL EY, Low-Level Control. By low-level control is meant: programs
that interact more-or-less directly with the effector mechanism, and that do
not engage in global planning or problem solving. Research on this topic is
producing a new and potentially important branch of classical automatic
control. Although little has been formalized to date, enough experience has
been acquired to permit the construction of interesting demonstrations.
Among the most impressive af these is an arn control program that can drive
the arm in partially constrained ways; for example, the arm can be made to
turn a crank by dynamically constraining the necessary degrees of freedom.

SUBGOAL E5. Hardware Development. The manipuletors available in 1966,
wnether based on prosthetic limbs or industrial put-and-take machinery, were
generally too primitive to be of long-term value for AI research. This
situation fostered a fairly significant hardware development effort that
produced a useful arm-hand device. Sinilarly, sensinz devices received sone
development efforts. Examples of this work are newly developed optical
range finders, and special tactile, force, and torque sensors.

Privileged Communication
OVERVIEW OF ARTIFICIAL INTELLIGENCE RESEARCH Appendix I

GOAL 3. INFORMATION PROCESSING PSYCHOLOGY: DEVELOPING DETAILED
SCIENTIFIC MODELS OF HUMAN SYMBOLIC PROCESSING BEHAVTOR.

Since its inception, one focus of AI research has been the study of the
symbol manipulation processes capable of explaining and predicting human behavior
in a wide range of cognitive tasks. As science, the endeavor is entirely
classical in intent and method, employing model construction and validation.
Empirical data from well-controlled laboratory experiments is obtained from
psychologists or generated by the researchers in their own laboratories.
Induction from this data leads to the formulation of a symbol-processing model
which purports to explain the observed phenomena. This model is given a precise
form as computer programs and data structures (since the conputer as a general
Symbol~processing device is capable of carrying out any precisely specified
symbol-manipulation process; this step is entirely analogous to the model-
implementation step taken by the physicist when he translates his physical model
into the form of a set of differential equations). A computer is then used to
generate the complex and remote consequences of the symbol-vrocessing postulates
of the model for the particular laboratory situations and stimuli being studied.
These consequences and predictions are tested against empirical data; differences
are noted and analyzed; the model is refined and run again; iterations continue

until a satisfactory state of agreement between model’s predictions and empirical
data is achieved,

From one point of view, the endeavor is to be seen as Theoretical
Psychology. From another point of view, it can be seen as a systematic attempt
by AI research to understand intellectual activity as it occurs in nature {i.e.,
in humans) so that artifacts capable of performing such intellectual activity can
be constructed upon the principles discovered. The interplay between these two
views has been very strong.

Information Processing Psychologists have usually chosen their problems in
areas that have been of "classical" coneern to Psychology, though some of these
areas have been reopened to serious investigation because of the successes of the
information processing approaches. The following are brief sketches of some
subgoals of the effort in Information Processing Psychology.

SUBGOAL 3A. Functional reasoning. Analysis and modeling has been done for
human behavior in solving logic problems, complex erypt-arithmetic puzzles,
and chess-play problems. The models, and the predictions derived from then,
are so detailed that no comoarison with previous work on the psychology of
problem solving is meaningful. The work is a scientific revolution, and has
had a great paradigmatic and methodological impact upon Psychology. The
principal innovators, Newell and Simon, have had their contributions
recognized by election to the National Academy of Science; Simon was awarded
the Distinguished Scientific Contribution Award of the Anerican Psychological
Association, more or less the "Nobel Prize" of Psychology.

ae

Privileged Communication 215 J. Lederberg
Appendix I OVERVIEW OF ARTIFICIAL INTELLIGENCE RESEARCH

SUBGOAL 3B. Rote Memory and Short-term Memory phenomena: Storage and
retrieval processes for short-term memory. Rote memorization effects.
Discrimination and association learning for verbal materials. These and
related phenomena of verbal learning and memory have been studied intensely by
experimental psychologists in this century. A few dozen solid empirical
generalizations are known. A set of closely related information processing
models is capable of explaining many of these (roughly speaking, 15-20 of the
"classical" phenomena).

SUBGOAL 3C. Long-term associative memory: Associative retrieval fron
associative memory nets of several hundred to a few thousand symbols.
Interaction of English sentence processing and memory. The symbolic
representation of knowledge (i.e., facts about the world) in memory. The work
is currently very active, highly promising, and is causing a mini-revolution
in thinking among psychologists who study memory.

SUBGOAL 3D. Pattern induction/concept formation. Induction of models of
pattern regularities in strings of symbols. Induction of the "senerating
rule" from the exhibition of instances of the rule.

SUBGOAL 3E. Phenomena of neurosis. The behavior studied is neurotic symbol-
processing behavior, viewed as processing distortions of otherwise "normal"
linguistic and problem-solving processes. A hizhly successful model of

paranoid behavior has been developed, ineorporating some English language
processing.

These examples are but pieces of a bigzer picture, which looks something
like this:

7. It is no surprise that Psychology nas been strongly affected by the
information processing concepts and tools of AI research since both sciences
are concerned with the study of cognition. The magnitude of the impact is
the big surprise. It is probably fair to say that the dominant paradigm
currently structuring Experimental Psychology in this country is the
information processing paradigm. Upon no other area of science has AI
research had such a strong impact.

2. The scientific study of human thought has been accelerated greatly during the
last fifteen years because of the AI impact. It is not much of an

overstatement to say that the AI impact has revitalized the study of thinking
by Psychology, making this scientific enterprise tractable, fruitful, and
respectable.

J. Lederberg 216 Privileged Communication
OVERVIEW OF ARTIFICIAL INTELLIGENCE RESBARCH Appendix I

VIEW OF THE FUTURE: What lies within a five year horizon?

An extrapolation of the research directions previously described into the
future faces at least two problems. First, there are the usual uncertainties
that loom because of unpredictable advances and wishful thinking. second, the
imposition by ARPA of research priorities upon the course of events that would
"normally" ensue will have a large effect. Tnus, the question of "what should
happen" is as big a question as “what will happen."

This exposition is made difficult by the fact that the structure of the
field, as outlined above in terms of Goals, will show strong confluences during
the future period. Any simple presentation goal-by-goal would be misleading, and

was not attempted. Instead, each identifiable focus is stated and then given an
extended discussion.

The main thrusts of the Artificial Intelligence community in the next five
year period will be:

1. Development of applications programs that represent and use knowledge of
carefully delimited portions of the real-world for high-performance problem

solving, hypothesis induction, and signal data interpretation.

The next period is likely to be a period of consolidation of AI’s
previous gains into meaningful real-world applications. High levels of
competence in the performance of difficult tasks will be the hallmark. In
addition to growing attitudes toward becoming more relevant, the AT
community’s current major interest in knowledge structuring and use will
naturally lead it to bodies of real-world knowledge tnat are rich in
Structure and challenges. An extrapolation indicates applications to
domains in science (much as the DENDRAL and MATHLAB programs were
developed); and in medicine (current activity includes programs that deal
with Infectious Diseases and with Glaucoma); perhaps more routine aspects
of architecture (e.g., space layout and design); perhaps design in
electronics (e.g., layout of IC and PC electronics, actual circuit design
to functional specs); management science applications (e.g., logistics
management and control, crew scheduling for aircraft fleets). The most
Significant application will be to computer science itself, namely the
automation of many programming functions (to be discussed later).
Application to some of the less routine aspects of office document
processing is a likely event (discussed later). With approvriate stimulus
from ARPA, or other service agencies, these application priorities could
be shifted toward defense problems, oarticularly those related to signal
processing (e.g., application to seismic or sonar signal interpretation).
In such applications, interpretation of what the signal means is made in
terms of knowledge about the Signal-generating source and the environment
in which the signal occurs. All of these applications will be
characterized by careful choice of domain, careful delimiting of the
extent of knowledge necessary to do the job, and close coupling with human
experts to gain the knowledge necessary. None of these programs will be
"general problem solvers" of the old genre. Characteristic of some of

Privileged Communication 217 J. Lederberg
2.

J. Lederberg 218

ix I OVERVIEW OF ARTIFICIAL INTELLIGENCE RESEARCH

these aoplications will be one-line interaction with human experts, not
only to "tune" the knowledge used by the program, but also to intervene in
decisions for which human expertise doninates that of the program, or
where the relevant knowledge nas not been made explicit and formalized for
computer processing,

The development, in particular, of that area of application involving the
Synthesis of computer programs (the so-called "automatic programming"
problem).

The particular application of AI techniques to the task of synthesizing
computer programs from imprecise and non-procedural descriptions of what a
user wants a computer to do for him is the AI problem area whose time has
come. This area will be the subject of a separate and detailed program
plan. It is an AI application of tremendous economic, and industrial
importance, since computer programming is today a major bottleneck in the
application of computers to technological and business problems. What is
worse, virtually no advances of substantial impact upon this problem have
been made in the last decade in other areas of computer science (with the
possible exception of the interactive editing, debugging, and running of
programs). The automatic programming problem is, furthermore, the
quintessential problem that fits the WHAT-TO-HOW characterization of the
nature of the science of Artificial Intellizenee. It is the meeting
ground of many of the tributaries of AI research: problem solving, theorem
proving, heuristic knowledge and search, understanding of English (perhaps
even speech), and advanced systems work. It is an ideal problem from the
viewpoint of knowledge-based systems--the main line of current AI
research. The essential activity in building such systems is the
extraction and formalization of knowledge of the specific task domain. In
the art of programming, computer scientists are their own best experts,
and for years have been engaged in formalizing what is known about
programming, mathematically and in other forms. Following this line of
reasoning, the programming task that may be best suited is systems
programming. An example of a specific systems programming task that may
be accomplishable within the period is: develooment of an automatic
programming system that will produce operating system code for a
minicomputer like the PDP11/45, in response to functional specifications
for instrument control and data-handling, wnera the snecs are given in
functional terms by a scientist putting together the instrument-computer
package, not his (until now inevitable) prozranner.

The extension of current ideas about the processing and understanding of
English to more extensive domains of discourse and with greater flexibility,
to the point of practical front-end processors for large applications
programs.

Privileged Communication
OVERVIEW OF ARTIFICIAL INTELLIGENCE RESEARCH Appendix I

In the coming period, programs for understanding English in limited
"universes of discourse" will achieve practicality, and will be made
available as the linguistic interaction vehicle in some of the larger AI
applications programs, e.g., the automatic programming systems mentioned
above. Since these applications programs will be domain-limited anyway,
it will not be an extraordinarily difficult task to construct for then
front-end processors that understand fnglish in that domain. Since
currently the field has only "demonstration programs" that exhibit
(limited) understanding of English, much more research will be undertaken
in these directions: examining how well current techniques extrapolate to
broader domains of knowledge; developing techniques for establishing
context of an interaction and maintaining that context throughout the
conversation; and extending methods for drawing inferences from the
continually updated context. Research on semantic theory, previously
mentioned in connection with representation of knowledge, will be applied
to specific problems of linguistic interaction involving actors, actions,
objects, and common-sense knowledge. The area of language understanding
is so rich in possibilities and implications that it is not unreasonable

to consider developing a separate program plan for it within the next two
years.

Initial exploration of office-work tasks as an area of development and
application; the careful choosing and shaping of specific tasks in this
enormous arena of human endeavor; and some limited applications progress on
these tasks.

The AI research community has been searching for problem domains of
Significance to science, technology, or industry that would provide an
integrating theme for the various subareas of AI work. These subareas
have a considerable coherence of concepts and techniques, but the
centripetal force of a real-world thene is necessary to make this
coherence a practical reality. Production assembly by combinations of
vision, manipulation and problem solving programs is an attempt to
establish such a theme. Increasingly the feeling is growing in the AI
community that the development of "intelligent assistant" programs for
ordinary office work is a.useful and important focus. There are two
reasons for this. First, much of current AI research fits the task area
well (e.g., semantic-net-memory structures, question answering programs,
natural language understanding, "intelligent assistant" interaction
programs, etc.). Second, the explosion of use of the ARPA network for
“office work" tasks quite apart from computation (uses such as message
processing, message and document filing, information retrieval from large
data bases, composing and editing of documents, ete.) provides an
excellent medium in which to do tne work. The AT community, perhaps with
a push from ARPA, has the capability to do significant work on the office
automation problem in the next period. A carefully thought-through
program plan will probably be the first output of tne field in this area
(should be organized and completed within the next two years), followed by
initial exploratory ventures along the lines laid down in the plan.

Privileged Communication 219 J. Lederberg
Appendix I OVERVIEW OF ARTIFICIAL INTELLIGENCE RESEARCH

J.

Again, as with all the knowledge-based systems of this decade, the
specific tasks worked upon will of necessity be carefully delimited. The
general “intelligent office assistant" is well beyond the horizon, but
specific assistant-programs for handling some of the office-work flow of
information on the ARPA network can be realized within five years.

Intensive developmental work on the speech-understanding problema.

Expansion of computer vision research to: knowledge-based program
organizations; development of a. repertoire of low-level perceptual operators
for color, range and texture, and exploitation of these modalities; first
practical applications of scene analysis to selected tasks in industrial and
biomedical settings; and use of interactive scene analysis for both research
and application purposes.

seene analysis programs consist of a combination of sensing-and-measuring
primitive perceptual operators (like line-finders) and higher-level
knowledge-based procedures (like line-proposers). Because of general
awareness of the limitations of current primitive operators (at least as
they are applied to monochrome pictures), the research will place
increased emphasis on the acquisition and low-level analysis of color and
range data. Higher level procedures will use knowledge of: three-
dimensional properties of objects other than polyhedral objects;
perceptual properties of objects; many varieties of contextual constraints
among objects; and properties of the orimitive operators (like
computational cost, reliability, and domain of applicability). Practical
applications will probably focus on industrial tasks like work-piece
identification and location, inspection, and manipulator control. The
scene analysis research issues in these applications may turn out to be
pedestrian, but concerns about cost, reliability, and reprogrammability
will become prominent. Biomedical scene analysis problems will continue
to stimulate research; application to medical mass-screening tasks may
occur. Interactive scene analysis will be an important focus. In research
settings, interactive scene analysis will be used to construct large
scene-analysis systems through the incremental accumulation of knowledge;
in application settings it will be used to achieve flexible scene analysis

systems that can be easily "re-programmed" by users who are not computer
selentists.

Lederberg 220 Privileged Communication
OVERVIEW OF ARTIFICIAL INTELLIGENCE RESEARCH Appendix I

7. &xpansion of arm-hand effector technology and associated progran control,

with some practical applications of simple forms of this technology in
industrial settings.

There will be considerable activity in the transfer of ARPA-initiated work
on effector control to industrial settings. Hardware realizations of a
rich variety of mechanical effectors, with their tactile, force, and
torque sensors, will appear. Visual feedback in controlling effectors
will be a feature of many of the applications. Basic research on the
hardware and software technology of effector control will continue, if
Support from ARPA or other agencies is forthcoming. More broadly-—based
research on effector control is likely to be stimulated by the appearance
of relatively inexpensive experimental hardware. Researchers who are
currently unable to develop one-of-a-kind devices because of their cost
will enter the field.

8. Expanded basic research on acquisition, deployment and representation of
knowledge to support knowledge-based systems development.

Though the main thrust of AI research is in the direction of knowledzge-
based programs, the fundamental research support for this thrust is
currently thin. This is a critical "bottleneck" area of the science,
Since (as was pointed out earlier) it is inconceivable that the AI field
Will proceed from one knowledge-based program to the next painstakingly
custom-crafting the knowledge/expertise necessary for high levels of
performance by the programs. In the next period, the following kinds of
fundamental explorations must be pursued and strongly encouraged:

a. é«iditional case-study programs of hypothesis discovery and theory
formation (i.e., induetion prograns) in domains of knowledge that are
reasonably rich and complex. It is essential for the science to see some
more examples that discover regularities in empirical data, and generalize
over these to form sets of rules that can explain the data and predict
future states. It is likely that only after more case-studies are
available will AI researchers be able to organize, unify and refine their
ideas concerning computer-assisted induction of knowledge.

b. Development of interactive interrogative techniques, eoupling a
program to a human expert, by means of which the program systematically
elicits from the expert particular facts, useful heuristics, and
generalizations (or models) in the domain of the human’s expertise.

Again, specific case-studies are desirable. Their development need not
await the arrival of English languase understandins programs to facilitate
the interaction and interrogation. (Stylized languages designed for the
specific case-study domains will serve for now.)

e. Exploration of a variety of methods for bringing together disparate
bodies of knowledge neld by a program to assist in the solution of

Privileged Communication 221 Jd. Lederberg
Appendix IL OVERVIEW OF ARTIFICIAL INTELLIGENCE RESEARCH

°

specific problems that the program is called upon to solve. The nature of
this problem was discussed earlier under Goal 2A. If there are to be a
number of Experts (i.e., specialized knowledge bases) interacting in the
solution of a problem, how should their interaction be arranged? Is the
an Executive Program "in charge" of sequencing the activity of the
Experts? If so, what is the nature of the Executive Progran’s knowledge
about each Expert, and the appropriateness of calling that Expert to
assist at a specific point in the process? Should the Experts be
relatively independent, each with its own situation-recognizer to trigzer
its activity? These particular questions are posed here not in an effort
to characterize the problem completely (or even adequately), but to give
the flavor of the experimental inquiry tnat needs to be pursued in the
coming period - a period in which major AI programming efforts will be
directed toward knowledge-based systems with multiple sources of
knowledge.

d. Theoretical and experimental studies of representation of knowledge.
This basic and difficult problem is not one that is likely to have a
"solution" in a five year period. Theoretical studies will continue to
search for a logical calculus in terns of which to formalize and store
knowledge in a fairly "natural" way and for logical processors that will
lism. Experimental studies will
attempt to deal with the usual nonhonoseneity of representation among
different bodies of knowledge directly, dy orogramming translations of
representations from one "natural" reoresentation to another as necessary

in those situations requiring comaunication between Experts for joint
problem-solving.

9. Continuing basic research on various mathematical-logical problems such as
formal models for heuristic seareh, theoren proving methods, and mathematical
theory of computation.

Because heuristic searen has been a central thene of AI problem solving
research, it is likely that attempts at mathematical formulation and
analysis of heuristic search methods will continue. No existing research
thrusts indicate that this work should have high priority at this time.
However, the situation is unstable in the sense that a few key results
(e.g., new theorems or, more likely, new formulations of heuristic search)
might cause a rush of activity along lines of formal analysis.

A similar situation attends theorem-proving research. There are currently
no critical ideas acting as a forcing function, but nonetheless the
problem appears to some scientists to be central for progress in the long
run. In their view, to state that mouter can be used as a "symbolic
inference engine” is equivalent to saying that it is a "logic engine"; and
what makes a "logic engine” turn over is a theorem prover over the domain
of some logical calculus. The searen fer appropriate logical calculi and
associated tneorem provers will therefore continue.

»

J. Lederberg 222 Privileged Communication
OVEAVIEW OF ARTIFICIAL INTELLIGENCE RESEARCH Appendix I

The work in mathematical theory of computation has been peripheral to the
AI mainstream, but recently has been gaining momentum and importance, and
will enter the mainstream as basic research for automatic programming
efforts. To write programs capable of synthesizing programs obviously
requires a thorough understanding of the nature of programs. One kind of
understanding is gained by formal description and mathematical analysis
(the kind of understanding we take so much for granted in some physical
Sciences and engineering), To the extent that useful formal descriptions
of how programs are put together and what programs do can be discovered;
and to the extent tnat powerful theorems can be proved within the
formalism; the work on mathematical theory of computation could aid
significantly in the practical work of constructing automatic program
synthesizers and verifiers. Thus, there are noteworthy "breakthrough"
possibilities in this area.

A prediction of the most likely course of events in these tasks of formal
analysis is that they will be low-key, low cost, high risk/high payoff.

10. Continuing research on modeling of human cognitive processes using
information processing techniques.

At the interface between AI and the psycholozsy of human perception and
thought, the research tempo has been increasing for some time. In the
coming period it is likely that new methodology, new conceptual insights,
and new models will have a continuing dramatic impact on Psychology. The
feedback to on-going AI research will continue to be important,
particularly in the areas of perception and memory. The principal
developments are likely to be these:

a. Methodological: analysis by program of the thinking-aloud protocols of
humans solving complex problems (i.e., "data reduction" that requires some
language understanding and complex inductive inference), resulting ina
speed-up in this critical empirical procedure of perhaps a factor of 100.
A typical complete protocol analysis of human data in a puzzle-solving
task currently takes, without computer assistance, 1990 hours.

b. Short-term memory. The processes of human short-term memory will be
so well modeled and understood as a result of research in this period that
the topic will cease to be of major theoretical interest to psychologists.

c. Long-term memory. A very good model of human long-term associative
menory will be developed. The orogran which realizes this model will be
given a great deal of "garden variety" knowledge of the everyday world, as
the basis for empirical testing. Such a model will undoubtedly prove to
be an important subsystem in larger programs that attempt language
understanding in contexts involving common-sense knowledge. Only the
beginnings of such a memory model exist today.

d. Visual perception. The most important impact of AT on Psychology in

Privileged Communication 223 J. Lederbers
Appendix [ OVERVIEW OF ARTIFICIAL INTELLIGENCE RESEARCH

the coming period may be the initial formulation on an information
processing theory of human visual perception of common 3-D forms, along
the lines of the visual processing concepts and operators developed by AT
vision research, Af vision research stands on the threshold of Psychology
awaiting an intellectual pusn like the one given to problem solving in
late Fifties. If the push is made, and is successful, it will noticably
dent the theory of visual perception in five years and totally capture it
within ten years.

J. Lederoerg

Nh
nN
42

Privileged Communication
AI HANDBOOK OUTLINE Appendix II

Appendix IT

Al HANDBOOK OUTLINE

NOTE:

The following material is a tentative outline of a handbook on artificial
intelligence planned for publication. It is not to be cited or quoted out of the
context of this report without the express permission of Professor F. A.
Feigenbaum of Stanford University. This handbook is intended for two kinds of
audience; computer science students interested in learning more about artificial
intelligence, and engineers in search of techniques and ideas that might prove
useful in applications programs. Articles in the first seven sections are
expected to apoear in the first volume to be published in preliminary form by
September 1977. The remaining articles are expected to appear in the second
volume to be published in preliminary form by June 1978.

The following is a brief checklist that was used to guide the computer
science students engaged in writing articles for the handbook. It is, of course,
only a suggested list.

i) Start with 1-2 paragraphs on the central idea or concept of the
article. Answer the question "what is the key idea?"

ii) Give a brief history of the invention of the idea, and its use in
A.I.

iii) Give a more detailed technical description of the idea, its
implementations in the past, and the results of any experiments with

it, Try to answer the question "How to do it?.

iv) take tentative conclusions about the utility and limitations of the
idea if appropriate.

v) Give a list of suitable references.

vi) Give a small set of pointers to related concepts (general/overview
articles, specific applications, etc.)

vii) When referring in the text of an article to a term which is the

Subject of another handbook article, surround the term by +°S; e.g.
+Production Systems+.

AI Handodook Articles

I. INTRODUCTION
A. Philosophy
B. Relationship to Society
C. History
D. Conferences and Publications
Privileged Communication 225 J. Lederberg