1983 AI SYMPOSIUM SUMMARY REPORT

• Document Type: CREST
• Collection: General CIA Records
• Document Number (FOIA) /ESDN (CREST): CIA-RDP86M00886R000500040010-5
• Release Decision: RIPPUB
• Original Classification: K
• Document Page Count: 57
• Document Creation Date: December 21, 2016
• Document Release Date: September 3, 2008
• Sequence Number: 10
• Publication Date: July 12, 1984
• Content Type: MEMO

Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
The Director of Central Intelligence
Intelligence Research &
Development Council
F7r(..rL
Executive Registry
84- qoo
STAT
12 July 1984
STAT
Executive Director of Central Intelligence
FROM Philip K. Eckman
Chairman, AI Steering Group
SUBJECT 1983 AI Symposium Summary Report
STAT
1. Enclosed is a summary report of our 1983 Symposium on
Intelligence Applications for Artificial Intelligence. The
unclassified report is intended to be a synopsis of the major themes
and issues which emerged during our Symposium last December.
2. In some cases, detailed viewgraphs from individual
presentations were provided in the Proceedings which were
distributed at the Symposium. Video tapes of selected ke speakers
are available through the CIA Self Study Center, A STAT
complete set of audio tapes was also recorded.
3. Over 600 people attended last year's Symposium. Plans are
underway to continue this annua event with a third Symposium in the
Spring of 1985 at a place and date to be determined.
Philip K. Eckman
C ys 'A) C I
~Dc(
STAT
Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
 Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
ARTIFICIAL
INTELLIGENCE
SYMPOSIUM
"INTELLIGENCE APPLICATIONS OF Al"
December 6, 7, & 8, 1983
A REPORT
Symposium Co-ordinator: SMART SYSTEMS TECHNOLOGY, Inc.
Specializing in Artificial Intelligence Education and System Implementation
Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
 Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
1. EXECUTIVE SUMMARY
1.1 SYMPOSIUM OBJECTIVES
1.2 MAJOR THEMES
1.3 CONCLUSIONS
2. DETAILED SUMMARIES OF SYMPOSIUM PRESENTATIONS
2.1 REQUIREMENTS ANALYSIS
Some points-of-view concerning the Intelligence
Community's requirements for computer-based
intelligent systems
2.2 OVERVIEW OF AI RESEARCH AND APPLICATIONS
History and scope of AI and summaries of some
current research trends
Some examples of current AI R&D projects in
Defense and Intelligence
Some examples illustrating how U.S. and Japanese
industry are structuring their investments in AI
2.5 INVESTMENT STRATEGIES BY DEFENSE
A Summary of DARPA's Strategic Computing Program
2.6 THE TACTICS OF AI TECHNOLOGY TRANSFER
The pragmatics of starting and maintaining AI R&D
facilities
APPENDICES
Appendix 1 AI Symposium Program, 1983
Appendix 2 AI Symposium Program, 1982
Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
 Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
1. EXECUTIVE SUMMARY
1.1 SYMPOSIUM OBJECTIVES
There is a strong current trend in both DoD and the
Intelligence Community, toward exploratory development of
computer-based intelligent systems for processing data. This
trend is a deliberate response to an urgent requirement:
senior defense and intelligence managers have long foreseen
that the continually increasing rates of acquisition of all-
source data can not be matched by any comparable increase in
the number of intelligence analysts available to process and
interpret the data. A central objective of Artificial
Intelligence is to enable computers to emulate the
intellectual functions that humans employ in analyzing and
interpreting data. For these reasons the Intelligence
Community is interested in assessing the potential of
Artificial Intelligence as a technological key to computer-
based intelligent systems for intelligence applications.
In December 1983 the Artificial Intelligence Steering Group
(AISG) of the Intelligence Research and Development Council
sponsored the second Annual Symposium on Intelligence
Applications of Al. The Symposium had several objectives:
o To provide a forum for the exchange of ideas and
experiences concerning the technology of artificial
intelligence (AI) and how it might be applied to
information collection, processing, and.analysis within
the Intelligence Community.
o To provide the audience of intelligence professionals
with a better understanding of the current state of the
art in AI, a more focused appreciation for where and
how AI can be applied to the intelligence business, and
an increased awareness of who the relevant AI players
are -- both inside and outside the Government.
The Symposia also provide a series of snapshots in time
against which the community can measure the pace of progress
in AI research and technology, and the rate of AI technology
transfer into intelligence applications.
The presentations in these Symposia were selected to provide
a balanced view of industry trends, academic research, and
intelligence applications. In order to allow in-depth
treatment of a few topics, the first two Symposia have
concentrated primarily on Computer-Assisted Image, Signal
and Speech Interpretation, Expert Systems, Intelligent Data
Bases and Overviews of Intelligence Requirements.
Deliberate omissions to date include robotics and AI
applications to resource allocation and planning.
Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
 Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
As intelligence applications of such technologies mature
they will be covered in future Symposia.
The presentations are organized in terms of six major
themes. These themes are concerned with the Intelligence
Community's requirements for computer-based intelligent
systems, the significance of AI research and technology for
realizing these systems, and the R&D investment strategies
and technology transfer tactics required to make it all
happen. These themes will significantly affect the
Community's future courses of action concerning AI research
and technology. In this Executive Summary we will present
the highlights of the speakers' remarks as they relate to
these major themes, indicating where the speakers provided
novel points-of-view or interesting technical detail. More
detailed summaries of the individual presentations are
presented in Sections 2.1 through 2.6 of this report.
1.,2 MAJOR THEMES
Theme #1: Requirements Analysis: Some points-of-view
concerning the Intelligence Community's requirements for
computer-based intelligent systems.
The speakers agreed that computer assistance is critical in
fusing, analyzing and interpreting vast quantities of all-
source data. The use of computer-based job performance aids
to increase analyst productivity was emphasized by several
speakers. Interesting detail was presented by Mr. David
McManis, NIO/W, concerning the urgency of the requirements
for natural language processing, for analyst aids in
accessing data bases, for better warning indicator
methodologies, and for unified concepts of information
handling. Emphatic expressions of the urgency of improved
information handling systems and analyst aids were made by
high-level managers in Defense and Intelligence.
Theme #2: Al Research and Applications: The history and
scope of AI, and summaries of some current research trends.
AI research and development spans a very wide spectrum of
activities with multiple objectives; primarily, the
practical objective of making machines smart in order to
make them more useful, and the scientific objective of
understanding the computational nature of human
intelligence.
Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
 Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
Today's Expert Systems are "idiot savants", demonstrably
useful in strictly circumscribed applications; DEC's XCON is
a landmark. Some Natural Language Understanding systems are
demonstrably useful; however, existing systems are far from
being able to recognize and respond to the user's goals as
distinguished from responding literally to queries. AIC's
INTELLECT is the most successful commercial Natural Language
product to date. Interesting detail provided by Dr. Patrick
Winston on MIT vision research from which new principles of
signal-to-symbol transformation are emerging. These new
"scale-space transform" techniques may eventually facilitate
building expert systems for signal understanding.
Interesting detail by Dr. Winston on Dr. Michalski's
research on "symbolic clustering" and on Dr. Winston's own
research on learning and reasoning by analogy. The long-
range payoff of such research may be realized in expert
systems which learn from experience, and which recognize
analogies between current and past situations and respond
accordingly. Additional interesting detail was provided by
Dr. Rodney Brooks concerning Stanford's approach to model-
based automated image interpretation using ACRONYM.
Partially successful initial results were reported; the
technology is still experimental.
Promising developments in the area of computer hardware and
programming environments for AI. Several Lisp machine
vendors: LMI, Symbolics, and Xerox, demonstrated very high-
quality AI workstations and Lisp programming environments.
DEC announced that Common Lisp and OPS5 will soon be
available as supported products for the VAX. Apollo and IBM
personal workstations also were shown as interesting low-
cost alternatives for limited AI applications. The cost of
AI workstations is still high reliability is improving, and
performance is generally quite good.
Theme #3: Intelligence Applications of Al: Some examples
of current AI R&D projects in Defense and Intelligence.
Seven AI research and development projects with clearly
defined defense or intelligence objectives were presented.
Three were directed towards automated signal processing and
interpretation. Each of these systems seeks to represent
human experts' knowledge about the engineering and
operational backgrounds of signals and to reason with this
knowledge in combination with traditional signal processing
algorithms.
Three of the projects were directed towards automated image
interpretation. All of these systems seek to represent
collateral information known to experts about objects and
events in the scene and to reason with this information in
combination with traditional image processing techniques.
Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
 Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
The seventh system seeks to reason with military experts'
knowledge about targets and battlefield operations in
combination with mathematical optimization algorithms to
achieve an objective of optimal resource allocation.
Only one of the seven systems described (TRW's signal sorter
tuner) is claimed to be functioning productively in an
operational environment. The projects are fairly recent
starts, and none as yet is close to the level of
accomplishment claimed for DEC's XCON. For the image and
signal-related systems the intrinsic difficulty of signal-
to-symbol transformation and the very broad background
knowledge required for these domains, are obstacles to
operational success.
Theme #4: Investment Strategies by Industry: Some examples
illustrating how U.S. and Japanese industry are structuring
their investments in AI.
Speakers from seven major U.S. corporations agreed that
industry should and will invest in AI because it is probably
important, useful, and profitable. Industry is making a
major thrust in expert systems for planning, diagnostics and
data interpretation. Interesting detail was provided on
DEC's corporate thrust towards building expert systems
modules for in-house applications in sales, manufacturing,
diagnostics, and system configuration. These modules will
eventually be integrated into a corporate-wide composite
system. Interesting detail on Westinghouse's corporate
strategy of "learning AI technology by doing it", training
its own AI engineers in-house, facilitating technology
transfer by establishing close ties with universities (CMU)
and by establishing in-house Centers-of-Excellence.
Interesting detail on treating situations where complex
combinatorics are required and where "humans find infinitely
many ways to screw things up." Emphatic testimonials by DEC
on expert systems' proven value to the corporation,
"comparable to the effects of the assembly line on Ford
Motor Company."
As to the Japanese ICOT Project: Key observation made by Dr.
John Alan Robinson that the complex and detailed project
plan published by ICOT obscures the simple central themes of
the project. The themes are: a commitment to make Logic
Programming the central language for symbolic computing and
a commitment to make the world's fastest (109 logical
inferences per second) logic computer by the 1990's. In Dr.
Robinson's view, this is rather like a dramatically enhanced
version of the U.S. Lisp Machine projects which also seek to
provide very high performance, low-cost symbolic processing
tools.
Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
 Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
Significant prediction: the Japanese may achieve their
engineering goals with the super-Logic Programming Machine,
but will find the AI sub-goals (Vision, Speech, Expert
Systems) just as difficult to accomplish as we have.
Theme #5: Investment Strategies by Defense:. A summary of
DARPA's Strategic Computing Program.
DARPA disclaimed that the Strategic Computing program is a
U.S. response to the Japanese ICOT program. Interesting
detail was provided by Dr. Robert Kahn, Director of DARPA's
Information Processing Techniques Office, on the program's
general goals. Like ICOT, these goals include symbolic
processing capabilities of 108 logical inferences per second
with generic AI software for vision, speech, and data base
applications. Unlike ICOT, there was no expressed commitment
to Logic Programming as a central theme.
Also, unlike ICOT, there is a specific focus on three major
military application areas: Navy battle management,
autonomous battlefield vehicles, and automated aircraft
cockpits. Contrast with ICOT: while the ICOT project is
dominated by a central theme, and implemented by a principal
team of forty AI researchers, the DARPA program is largely
directed toward facilitating the efforts of unspecified and
independent research groups from all over the U.S.. These
independent groups will have access to computing hardware
and facilities for designing and manufacturing innovative
microprocessors and multi-processors.
Theme #6: The Tactics of Al Technology Transfer: The
pragmatics of starting and maintaining AI R&D facilities.
The observations of six experienced R&D managers concerning
in-house AI Centers as a mechanism for facilitating AI
technology transfer may be summarized as follows:
o In-house AI facilities are a very good mechanism for
developing AI capabilities within an Agency,
o The facilities' research programs should be closely
coupled to the Agency's mission,
o The scarcity of qualified AI personnel is a critical
problem . . . you have to train some of your own,
o Long-term investment commitment by high-level
management is essential for stability and success of
the AI facility,
o Computing equipment and software should be selected to
facilitate sharing software with universities and other
research groups.
Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
 Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
1.3 CONCLUSIONS
The national technology base for building intelligent
systems is substantial and rapidly growing. Practically
every major high technology company in the U.S. is engaged
in exploratory development projects related to computer-
based intelligent systems, both for internal use and as
products. Growth in the national AI technology base is
spurred by continuing price/performance improvements in
computer hardware, by DoD spending on intelligent systems
R&D, by consumer expectations for "smart" products, by
several AI technology successes in knowledge-based systems
and database interfaces, and by the perceived need to
compete with Japanese technology thrusts. It is virtually
certain that AI technology growth will continue and will
accelerate. Generally, private industry should not expect
immediate major payoffs from its investments in AI.
The Intelligence Community has urgent and specific
requirements for computer-based intelligent systems. The
massive data-acquisition capabilities of our collection
systems exceed, by at least a factor of ten, the
intelligence processing capabilities of the severely
limited number of experienced intelligence analysts
available. Intelligence Community working groups have
identified specific problem areas where computer-based job
performance aids for intelligence analysts would have high
payoff. These include: data base access and automated
data-base construction; foreign language translation; image
and signal analysis; intelligence resource allocation and
planning; capturing unique knowledge and skills in
knowledge-based systems; and aids to data fusion and
interpretation. AI R&D is addressing each of these areas,
but no generic solution has been found for any of these
applications. The most significant progress to date is in
data base access and in knowledge-based systems, where
several commercially viable systems have been developed by
industrial firms focusing on very specific problems.
The Intelligence Community's current program in AI
technology applications has not reached critical mass.
Several leading corporations including Westinghouse and DEC
are exerting, on their own behalf, AI efforts substantially
larger than the Intelligence Community's. Factors that
senior corporate managers judge to be essential to
successful major AI programs include: firm, long-term
management commitment and project goals that are both
technically achievable and significant to the corporation's
business.
Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
 Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
STAT
Some corporations address the current scarcity of
experienced AI technologists through staff training
programs, in-house centers of AI expertise, and links with
universities. By contrast, the Intelligence Community is
currently pursuing perhaps a dozen relevant but relatively
small exploratory development efforts in AI systems. These
are primarily conducted by outside contractors. An AI
orientation program for Intelligence Community managers has
recently been initiated. While these steps are in the right
direction, the overall effort still lags far behind critical
mass when measured against the magnitude and urgency of the
Community's requirement to reduce its information-processing
overload. We should be preparing system specifications,
initiating substantial development programs, and training
technical personnel now if we expect to see computer-based
intelligence in our information processing systems by the
mid-to-late 1990s.
These remarks conclude our, Executive Summary of the December
1983 Symposium on Intelligence Applications of AI. In
Sections 2.1 through 2.6 following, more detailed summaries
of the major presentations are provided.
For access to the original data the reader is referred to
the Symposium Proceedings and to video and audio tapes made
of the symposium. available through the Symposium Chairman
Office of Research and Development,
Central Intelligence Agency, Washington, D.C. 20505
STAT
Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
 Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
2. DETAILED SUMMARIES OF SYMPOSIUM PRESENTATIONS
Artificial Intelligence is the most rapidly expanding
research and development activity in applied computer
science today. Specialized AI computers and AI software
systems have migrated from university research laboratories
into exploratory development projects in several government
laboratories and in practically every major high-technology
industry in the U.S. The presentations in this Symposium
were selected to present a balanced overview of current
industry trends, academic research, and government programs
related to intelligence applications. In the following
section of the report, the presentations are organized in
terms of six general themes. In the spirit of presenting
objective data for Intelligence Community managers and
decision makers, we have attempted to make this report a
reasonably neutral account of what the speakers actually
said.
Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
 Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
2.1 REQUIREMENTS ANALYSIS: Some points-of-view concerning
the Intelligence Community's requirements for computer-based
intelligent systems.
2.1.1 Speaker: Dr. Philip K. Eckman
Chairman, Intelligence Community AI
Steering Group
Topic: Symposium Objectives
Summary: Dr. Eckman foresees that in the coming
decade, the amount of all-source information will increase
tenfold and the number of analysts only by 20%. To
paraphrase Von Neumann: it's an insult to a man to have him
do a job that a machine can do as well. The issue, in Dr.
Eckman's view, is how to make people more productive and
reserve for them tasks that are uniquely human.
.2.1.2 Speaker: Dr. Richard D. DeLauer,
Under Secretary of Defense for
Research and Engineering
Topic: AI and Defense
Summary: Dr. DeLauer perceives AI and
supercomputers as vital to the defense and intelligence
communities. Supercomputers and high speed processing can
help DoD and the Intelligence Community overcome our
adversaries' numerical advantage by advancing our national
capabilities in the areas of training, information handling,
heads-up displays for cockpits, fusion of sensor data, and
battle management. In Dr. DeLauer's view, we need to
rapidly fuse, analyze and distribute large amounts of
intelligence data; AI will be important to us by helping
make this less labor intensive. As a direct result of this
Symposium, Dr. DeLauer and his staff intend to assess and
collate ideas received from the Symposium audience and will
use these ideas as a source of guidance on possible new ways
of using AI and directing further research in AI.
2.1.3 Speaker: David Y. McManis
National Intelligence Officer for Warning
Subject: Intelligence Requirements
Summary: What is missing to date, in Mr. McManis'
view, is a unified concept of information handling, and
that's really what AI should be about.
Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
 Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
The NIO/W is concerned mainly with warning, in the extended
sense . . . strategic and tactical warning, warning of
"suprise attack" of the Pearl Harbor variety, and with other
forms of threat, such as technology breakthroughs or
political and economic instabilities. Mr. McManis observed
that we do a generally good job in current and long-term
intelligence; the missing area is the six-month time frame
for warning and forecasting. Several areas that need to be
addressed include:
o Automated data bases, such as bibliographic,
--------- ---- b--
biographic, order-of-a -ttle, and geographic data bases.
There is a need to get the analyst out of the loop, and
to automate the process of building data bases. ELINT
data is well suited to this approach. With respect to
open source text, there is an urgent need for natural
language processing, even if only 50% of the job gets
automated, as with machine translation.
o Analytic access: in Mr. McManis' view, it is incredible
to watch senior analysts having to weave their way
through the labyrinths of our data base systems.
Better terminals and better tools for manipulating data
should be developed which remove the barriers between
the analyst and the data.
o Indicator methodologies for recognizing swing events
that flag harmful situations are urgently required.
Mr. McManis suggested that ORD might provide some
assistance in applying expert systems against the
Indicator methodology problem. Some "warning models"
are being built to help the NIO/W understand what the
key indicators are, and how to collect against them.
NIO/W is interested in imagery targets with low
activity levels but high indicator value; comparing
yesterday's pictures with today's may prove vital.
o Presenting information to decision makers so as to
answer their questions quickly: Here a tremendous
amount of help is needed in providing multiple
alternative views of data in response to different
types of questions and different levels of security
classification.
Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
 Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
2.1.4 Speaker: Mr. John N. McMahon
Deputy Director of Central Intelligence
Topic: AI and Intelligence
Summary: Mr. McMahon recalled that twenty-five
years ago the Intelligence Community talked about the
information explosion and automated information processing,
but history has shown the Community is smarter at building
big collection systems than at processing information. AI
can make significant contributions in machine translation
and photo interpretation. The Community's experiences with
AI to date have been somewhat "spotty"; the results have
been marginal. Nevertheless, in Mr. McMahon's view the
capability is out there, but we haven't been smart enough to
harness it.
Threats to U.S. national security include such diverse
phenomena as Soviet technological advances in lasers and
particle beams; Soviet military buildups; the cascading
effect of the Third World's $625 billion dollar debt; the
growth of economic competition from Europe and Japan.
Intelligence is required to keep track of all this, and in
Mr. McMahon's view, the AI Community can contribute to U.S.
National Security by helping the Intelligence Community
handle such problems. Mr. McMahon observed that: "the
Intelligence Community can't use poverty as an excuse; since
we've had a 75% budget increase in the last three years.
Let's do something with it!"
Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
 Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
2.2 OVERVIEW OF AI RESEARCH AND APPLICATIONS: History and
scope of AI and summaries of some current research trends.
2.2.1 Speaker: Dr. Patrick H. Winston
Director, Artificial Intelligence
Laboratories, MIT
Topic: Overview of Artificial Intelligence
Research and Applications
Summary: In Dr. Winston's view, AI is defined by its
objectives, rather than by its techniques. The objectives
are: making machines smart in order to make them more
useful, and understanding the nature of intelligence . . .
which Dr. Winston referred to ironically as the "Nobel
Laureate motive".
When we understand something thoroughly it no longer seems
mysterious; when we understand how AI programs work, they
may cease to appear "intelligent". Dr. Winston related an
anecdote concerning Dr. Slagle's first symbolic integration
program, a precursor to the MACSYMA system developed at MIT
in the 1960s. Dr. Winston taught the program in his AI
course at MIT; after the class one student exclaimed: " . .
that program isn't really intelligent . . . it integrates
the same way I do."
Dr. Winston summarized current research and applications in
five AI subdomains: Expert Systems, Natural Language
Processing, Vision, Signal Understanding, and Learning.
Expert Systems are "applied logic". Three major categories
include:
o Expert systems for identifying things; the MYCIN
diagnostic expert system is an example,
o Expert systems for configuring things; DEC's XCON
system is an example,
o Expert systems for interpreting signals; the
Schlumberger dipmeter advisor is an example.
Dr. Winston expressed suprise that so much has been
accomplished with such simple tools as rule-based systems.
Today's expert systems are "Idiot Savants", brilliant in
narrow domains, otherwise helpless. The next generation of
systems must be able to shift among alternative points-of-
view; must degrade gracefully as the knowledge base is
reduced; must be able to break their own rules, build models,
and most significantly learn from experience.
Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
 Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
A good domain for Expert Systems is one where: specialized
knowledge is required as distinct from generalized "common
sense"; where humans require an hour or so to solve the
problems; where an expert is committed to the project; and
where the subject matter is systematized to the extent that
there are books or manuals on it.
Natural Language Understanding requires a broad spectrum of
varieties of intelligence and knowledge. We now have
operational systems that do useful things, providing access
to data bases in an easy, natural way. We are a long way
from having systems that understand the speaker's intent and
frame-of-reference sufficiently to respond to queries in a
helpful rather than a literal fashion. As with Expert
Systems, Dr. Winston expressed surprise that so much has
been accomplished with such simple tools as semantic
grammars and frames. INTELLECT was singled out as the most
successful commercial natural language product. Frame-based
systems have been applied to such tasks as skimming
newspapers . an application that might be of great value
to the Intelligence Community. Because the narrative may
misfit an inappropriate frame, frame-based systems can make
hilarious mistakes and must be used with great care.
Vision and Signal Understanding: Dr. Winston observed that
work in vision has led to a principle for signal-to-symbol
transformation that may enhance expert systems for signal
understanding. The spatial second dervatives of
deliberately blurred images cross zero at points in the
image where humans tend to perceive edges. By
parameterizing the degree of blur and plotting the loci of
such zero crossings the image is transformed into "blur
versus zero crossing" space or scale space in Dr. Winston's
terminology. This leads to a natural segmentation of images
and signals. The transformation is invertible, and with
this transformation, signals can be divided into pieces,
much as the human eye would divide them up.
Learning: Dr. Winston illustrated Dr. Michalski's program
for separating diseased from non-diseased soy-bean plants,
"discovering" the combination of properties that
discriminates among classes. The program searches a vast
tree of possible combinations of properties, using induction
heuristics . . . a sort of "symbolic cluster-analysis
algorithm", and in this sense it accomplishes a symbolic
learning task. Dr. Winston proceeded to describe his
research in the area of "learning by analogy". One of his
programs takes as input very simple stories . a precis
of Macbeth was used as an example . . . and outputs a
semantic network relating actions by the play's principal
characters. The arcs in this net represent causal relations.
The nodes can then be generalized, and the relationships
treated as general principles.
Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
 Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
When other new stories are mapped onto the resulting
structure the principles may become hypotheses about the new
situations, arrived at "by analogy" with the initial
structure. For example, similar motives might be imputed to
protagonists in sufficiently similar narratives.
Dr. Winston views this research in analogical reasoning as a
very tentative but promising first step into a new area . .
"like being at Kitty Hawk in 1903, planning the program that
will lead to the 747". He has extended the approach from
stories to "object form and function," and is writing
programs that may enable robots to infer useful properties
of an object . . . such as "being liftable" . . . from
analogies and formal resemblances to other known objects.
Computation is a major impediment to extending such toy
learning systems to real applications. The sorts of five-
order-of-magnitude advances which are the goals of the DARPA
Supercomputing program are required here.
In summary, Dr. Winston observed, progress in AI is a locus
of points in "Applications versus Knowledge space," as shown
below.
Natural Language
I
F-Ru e-Based Systems
Learning
Signal-to-Symbol Transformation
Most of AI is still at or near the first riser of this step-
like curve. Our attitude towards AI, in Dr. Winston's view,
should be one of "restrained exuberance".
2.2.2 Speaker: Dr. Rodney Brooks
Assistant Professor
Computer Science Department
Stanford University
Topic: Overview of AI Applications in
Computer Vision and Image
Understanding
Summary: Dr. Brooks addressed the problem of
automated identification of objects in an image.
Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
 Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
There are two traditional approaches: image based, i.e.
processing and segmenting the image, working in the
direction of inferring the objects that might have given
rise to the observed image data; or model based, inferring
from an a-priori object model the image that it might have
produced, and comparing the inferred image with the observed
image data.
The ACRONYM system developed by Dr. Brooks and Dr. Binford
at Stanford University, represents a model-based approach to
image understanding. The primitive components from which
complex object models are built are generalized cylinders.
These parametrically defined tubular structures are combined
to form more complex objects . . . aircraft models, in the
example under discussion.
The parameters of a camera model, together with the
parameters defining the length, cross-section and curvature
of the generalized cylinders, suffice to compute classes of
predicted images.
Dr. Brooks experimented with digitized aerial photographs of
aircraft in commercial airports. Generalized cylinders were
fit against the edge-enhanced imagery, and the parameters of
the cylinders were checked for consistency against the
camera model and against object models for known aircraft
types. The experiment met with limited success. Failures
to recognize some objects resulted in part from failures in
edge detection and in part from the computational
impracticability of applying rule-based consistency checks
against the entire image, rather than against pieces of it.
In Dr. Brooks' very forthright summary he asserted that this
is still an experimental technology; it's hard to make the
computations parallel, and the performance currently is
poor.
2.2.3
Speaker:
Dr. Raj Reddy
Topic:
Director, The Robotics Institute
Carnegie-Mellon University
Overview of AI Applications in
Summary:
Robotics and Speech Understanding
No matter what area of AI we're working
in, Dr.
Reddy observed, knowledge representation is the key
ingredient. For example, Herb Simon once undertook as an
experiment the task of encoding the knowledge in one chapter
of a physics text (on statics). After six months he had
developed twelve condition-action rules which sufficed to
solve nineteen out of twenty-two of the problems at the end
of the chapter. What we still don't know how to do in AI is
to "give the computer this textbook and tell the machine to
distill the knowledge and solve the problems."
Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
 Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
Consider the variety of types of human knowledge. A small
fraction is algorithmic. Perhaps ninety-five percent is
informal, derived from examples. In robotics applications
the extreme difficulty we find in constructing a machine
like the hexapod walker, for example, underscores the need
for "understanding" how we walk, how we see, or how we
process speech.
In general, Dr. Reddy observed, we still don't fully
understand what knowledge is involved in these actions, how
it is represented, or how the knowledge is used the
organism to solve problems.
Dr. Reddy proceeded to describe the CMU Robotics Institute.
The staff numbers approximately one hundred and fifty, with
approximately fifty PhDs and sixty university students.
Westinghouse is a major industrial contributor to the
Institute. The major research themes are:
o Manufacturing Facilities of the Future,
o Operations in Hazardous Environments.
In the first area, knowledge based simulation looks
promising. AI-based simulators for training Westinghouse
production-plant operators cost approximately one million
dollars to build and are saving the company approximately
five million dollars per year. Other general problem areas
include: how to build flexibility into a production
facility; resource allocation . . . what resources are
required by a facility for a product; allocation of control
what operations should be autonomous, and where
should human supervision be employed. In general, Professor
John McDermott of CMU observes that "white-collar robotics"
is easier than "blue-collar robotics".
In the Hazardous Operations area, Sutherland and Moravec are
designing a six-legged walker and a three-wheel autonomous
rover. The control problem for the six-legged walker is
still not completely solved. Moravec's stereo vision
comparator for robotic vehicles highlights requirements for
faster computation: as Dr. Reddy remarks, it "takes one
step, then thirty minutes to think, takes one step, then
thirty minutes to think . . ..If
2.2.4 Speaker: Dr. Gian-Carlo Rota
Professor of Applied Mathematics
and Philosophy, MIT
Topic: Observations and Speculations on
the Effect of the Computer on Scientists
Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
 Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
Summary: Dr. Rota observed that in computer chess,
the winning programs have proved to be those that employ
brute force search rather than relying on ingenious
heuristics. "It's a sad commentary on the human condition
that reliability wins over genius!" Dr. Rota proceeded to
make three principal observations regarding AI:
o The simple-minded machine is most likely to succeed,
e.g. computer chess,
o There is no reliable way of telling beforehand whether
an AI program is easy or hard . . . there's no better
test than trying to build things; that's when we find
out whether our "common-sense" descriptions match
reality,
o Advances in AI have come from the "hard" rather than
the soft sciences.
As to international competition, Dr. Rota observed that in
the U.S., the free exchange of information among scientists,
and our superior university system, give our research
establishment an advantage over the Russians and the
Japanese. A serious national problem, however, is that we
allow our best technology to "drain" overseas. Dr. Rota
expects Hitachi to announce early in 1984 a machine with
five times the capability of a CRAY1. Fujitsu will announce
a machine with ten times CRAY capability. These rapid
advances by the Japanese computer industry are due to U.S.
failure to protect this technology.
In Dr. Rota's view, two new frontiers for AI technology are:
o transition from digital to holographic computation,
o the unexpected relevance of the theory of random nets
and non-deterministic programming to AI computation.
2.2.5 Speakers: Dr. John Vittal - Xerox
Mr. William Kaiser - Apollo
Mr. Steve Lazerowich - Symbolics
Mr. Robert Abramson - DEC
Dr. David Yun - E-Systems
Dr. Frank Spitznogle - LMI
Topic: Industry Panel: Description and Schedule
of AI Demonstrations
Several computer manufacturers made AI processors available
during the three days of the Symposium for demonstrations of
Al tutorial programs and working AI systems. These were
available during the mid-day breaks and following the
afternoon sessions in the Tunnel adjacent to the main
Auditorium at CIA Headquarters.
Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
 Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
2.2.5.1 Xerox Summa: Dr. John Vittal described
Xerox AI program development tools . . . primarily Interlisp-D and
Smalltalk, and the Xerox personal workstations with high-
resolution graphics. The emphasis is on programmer
productivity enhancement through interactive programming.
The Xerox demonstrations include:
o Interlisp D, the Lisp programming environment,
o LOOPS (object-oriented programming system with
constraints facilities),
o RABBIT (a mechanism for accessing databases),
o TRILIUM (a knowledge-based system for designing copier
interfaces),
o FORMS (a system to facilitate designing forms),
o PAPERWORKS (a system to facilitate annotating reports),
o MAPS (a system for annotating maps),
o SMALLTALK (an object-oriented program development environment),
o ANALYST (an analyst aid, facilitating interaction with
maps, text, and mail)
2.2.5.2 Apollo Summary: Mr. William Kaiser described
the workstation architecture: 32 bit VLSI processor and bit-
map displays, with individual workstations interfaced
through a local area network. Their major market is
computer aided design; the Apollo supports FORTRAN, PASCAL,
and C, color graphics, and Bell and Berkeley Unix.
o T, (a dialect of Lisp developed at Yale University)
o Portable Standard Lisp
o DUCK (the Smart Systems Technology logic programming
system for developing expert systems)
o KES (the Software A&E expert system application
generator)
o SIL (the SILMA programming language, for high-speed
graphics applications)
2.2.5.3 Symbolics Summa: Mr. Steven Lazerowich
described the Symbolics demonstration programs:
o Thoughtsticker (the Pangaro, Inc. tool for
representing alternative views of concepts)
o SARSYS (A TASC, Inc. system for radar image
interpretation)
o KNOBS (the MITRE experimental system for air mission
planning)
o INCA (VERAC's DARPA-sponsored data fusion project)
o SAGE (Symbolics' document support system)
Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
 Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
2.2.5.4 DEC Summar : Mr. Robert Abramson described the
DEC AI Technology Center in Hudson, Mass. DEC is a user and
developer of AI systems, such as XCON, and intends to be a
vendor of systems such as COMMON LISP and OPS under VMS.
DEC demonstrations include:
o XSEL (an expert system for guiding DEC sales
representatives in selecting components; XSEL calls
XCON for configuration assistance)
o OPS (Production System)
o COMMON LISP (DEC's implementation)
2.2.5.5 E-Systems Summary: Dr. David Yun of
Southern Methodist University, consultant to E-Systems,
described a low-cost, high-availability knowledge-base
system development environment on the IBM personal computer.
The hardware/software combination cost is approximately
three thousand dollars.
2.2.5.6 LMI Summary: Dr. Frank Spitznogle described
several features of the Lambda machine and the systems to
be demonstrated. The systems include:
o Lisp Machine programming environment on LMI's Lambda
machine,
o PROLOG (compiled into Lisp Machine lisp),
o TI Natural Language System,
The system architecture is based on a high-speed bus (the
NU-bus). The Lisp processor is interfaced to this bus; a
68010-based UNIX processor is also interfaced to the bus,
allowing either UNIX or Lisp programming.
The LMI implementation of PROLOG is expected to run
extremely fast and will be integrated into the Lisp
programming environment. A natural language parser and a
"smart arithmetic" demonstration were used to present PROLOG
concepts.
Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
 Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
2.3 INTELLIGENCE APPLICATIONS OF AI: Some examples of
current AI R&D projects in Defense and Intelligence.
These presentations were made during the second day of the
Symposium, in sessions classified at the Secret level. In
order to facilitate distribution, this report is
to be kept unclassified. This section summarizes the
unclassified aspects of the presentations.
2.3.1 Speaker: Dr. James Slagle
Senior Scientist, Navy Center for Applied
Research in Artificial Intelligence
Topic: BATTLE, an Expert Advisor for Weapons
Allocation
Summary: At the Navy Center for Applied Research
in Artificial Intelligence (NCARAI), several research and
exploratory development AI projects are underway:
o Expert Systems in Combat Management,
o. Expert Systems for equipment troubleshooting,
o Target Classification, using ISAR data on vessels,
o Natural Language Processing, applied to Navy message
automation,
o Multi-sensor fusion,
o Adaptive control.
Dr. Slagle described the BATTLE system in the context of
combat management. BATTLE has been implemented at NRL by
Jim Slagle and several colleagues at NCARAI. The objective
of the system is to provide improved weapons allocation
plans for Marine Corps Artillery and Air Support. Data
supplied by forward observers allows real-time updating of
the plans in response to target damage reports. The system
performs two sorts of computation: analysis of the
effectiveness of weapon-target combination and complete
multi-weapon, multi-target allocation plans.
The weapons effectiveness calculation is implemented by a
computation network, which is a generalization of
PROSPECTOR. Rules are specified by military experts and
stored in a data base. The allocation plan is computed by
an algorithm which searches efficiently through the space of
possible allocation plans for optimal, or near-optimal,
allocations which maximize damage to targets. BATTLE has
been tested by a Marine Corps artillery expert and was
judged to produce valid allocation plans for real problems
in reasonable. time. Other problems, such as assigning
individuals to jobs in an organization or parceling out
computational tasks in a multi-processing environment, may
be amenable to similar approaches.
Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
 Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
2.3.2 Speaker: Dr. Gerald A. Wilson and
Dr. Robert Drazovich
Advanced Information and Decision
Systems, Senior Computer Scientist
Topic: Computer-Based Assistants for Science
and Technology Analysis
Summary: Dr. Wilson described the prototype Expert
Assistant for Science and Technology Analysis (ASTA). The
objective of the ASTA project is to develop an interactive
computer-based "intelligent assistant" to facilitate
analysis of shipboard radars. The ASTA computer terminal
presents the user with a sequence of frames requesting
parameters on the radar under analysis. The user extracts
these parameters from photographic, radar intercept or other
available data, and enters the required information into the
frame. ASTA proceeds to "reason" with this data, using
facts and hypotheses about radars stored in its knowledge
base. The system may prompt the user for additional data as
required for its analysis.
ASTA has four major components: the dialog manager which
controls the user/system interface; the activit su~ort
mana9Ler, which maintains dynamic data bases of user
hypotheses and general radar knowledge; the information
manager which controls access to local and external static
databases; and the support tool manager which accesses radar
analysis algorithms. The rule-based sytem architecture
underlying ASTA is MRS, developed by Dr. Genesereth at
Stanford University.
ASTA is currently under development. The system handles
only a fraction of the information which could be provided
and only "knows about" a small number of radars. However,
Dr. Wilson observed that the results are quite promising and
that experienced radar system analysts are pleased with the
way in which ASTA helps them to record information in a
structured way.
Dr. Drazovich described the Advanced Digital Radar Image
Exploitation System (ADRIES). The goal of the ADRIES
project is to develop an interactive computer-based image
exploitation workstation to facilitate analysis of synthetic
aperture radar (SAR) imagery. This development is in the
context of a DARPA program involving eight companies. A
final demonstration is scheduled for the Summer of 1986.
The workstation is intended to assist the user through:
o "smart" techniques for automating some SAR image
analysis tasks,
o detecting and classifying some tactical targets
(missile sites, regiments, etc.),
o supporting the use of collateral information on the SAR
image to narrow the area search.
Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
 Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
The ADRIES workstation will enable the analyst to display
SAR imagery at several resolutions. Collateral information
will be used by the system to narrow the area search for
specific targets to "most likely" local regions which will
then be magnified and enhanced. The collateral information
includes current battlefield intelligence, enemy order-of-
battle, and operating characteristics of the targets. The
development system architecture consists of a VAX 750
mainframe, VICOM image processor, and a Symbolics 3600 Lisp
Machine, all interfaced through Ethernet.
2.3.3 Speaker: Mr. Mark Williams
Senior Staff Engineer,
ESL, Inc.
Topic: AI in Signal Processing
Summary: Mr. Williams described a research project
underway at ESL, exploring the application of expert systems
concepts and techniques to signal collection and analysis.
In traditional signal processing, Mr. Williams observed, the
"front-end processor" extracts signal features and makes
statistically based decisions using pre-determined
algorithms. The objective of the expert systems approach is
to increase the flexibility of the signal processor in
dealing with noisy or non-standard situations by enabling
symbolic reasoning about the signal, the environment, the
available collection of alternative processing techniques
much as a human expert might. Mr. Williams
represented the idealized "intelligent signal processing
paradigm" in the schematic shown below.
INTELLIGENT SIGNAL PROCESSING PARADIGM
SIGNAL
INPUTS
INFORMATION
AND DATA
REPRESENTATION
- PROCESSED
OUTPUTS
? PRESENT SITUATION
? BACKGROUND
INFORMATION
SIGNAL
PROCESSING
RESOURCES
PROCESSING
AND DECISION
MAKING
(INFERENCE ENGINE)
KNOWLEDGE
BASE
? SP EXPERTISE
? TOOL USAGE EXPERTISE
? APPLICATION DOMAIN
EXPERTISE
Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
 Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
ESL is currently experimenting with these AI approaches to
signal classification. The development system arcahi Xerox
consists of a signal digitizer, VAX mainframe,
1100 (Dolphin) Lisp Machine connected via Ethernet.
Close
2.3.4 S Baker: Programa Manager
Hughes Aircraft Company
Topic: Image understanding
Summary: Dr. Close described the overall objective
of the Hughes program as a demonstration of the feasibility
of automated port monitoring using Phase
y.
demonimagery
Eighteen months into the project, the beot
was conducted on a general purpose computer, using
available software developed up to that time by the image
understanding research community.
The ACRONYM system was selected for the project. ACRONYM is
described in greater detail in Dr. Rodney Brooks'
presentation on model-based image interpretation (Section
2.2 of this report). The initial experiments were rvoaluablee p
o and instructive. On the epositive si: d differentimageseof the same
systems successfully align ~ probable interest,
scene, successfully detected objects of pand correctly identified the approach to image interpretation.
validating the model-based
On the negative side: the ACRONYM system was designed as an
image-understanding research tool and was found to have a
number of specific deficiencies for this particular
application; the available object models were too limited,
and the system's handling of rules and data structures was
less flexible than required.
For the Phase 2 technology development foll his collea k (an and gus
on-going two-year effort), Dr. intend to incorporate a detailed camera mod
control; the
system; to add knowledge-based planning a
incorporate three-dimensional computer graphics object
modeling and image processing algorithms to f cilitate the
image intensity and texture processing juged to
necessary for object detection and classification in the
port monitoring application.
Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
 Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
2.3.5 S taker: Dr. Dick Kruger
Senior Scientist
Science Applications, Inc.
12 2-1-2: AI Applications in Synthetic
Aperture Radar Image Interpretation
Summary: Dr. Kruger presented an outline of SAI's
recent research and development initiative in rule-based
synthetic aperture radar (SAR) image exploitation. The
objective of the effort is to use terrain knowledge and
collateral information to narrow the area search for targets
in SAR imagery. The underlying data base consists of DMA
terrain maps of the Fulda Gap area, combined with SAR
imagery of the same area taken during ongoing military
exercises (REFORGER, 1981). Mobility data is derived from
the terrain maps. Rules relating vehicle characteristics to
terrain mobility are used to characterize certain geographic
regions as "denied", thereby directing the search for
targets to more likely areas. Initial results appeared
promising and the rule set is currently being extended. The
OPS-5 production system was used for the underlying rule-
based-systems architecture.
2.3.6 S taker: Mr. John C. Kelly
Senior Staff Engineer,
TRW Defense Systems Group
To ic: An Operational Artificial Field
Engineer for Tuning a Signal Sorter
Summary: Mr. Kelly's organization collects and
processes signals which are sorted into "bins" based on the
features of their time-versus-frequency plots. In the
traditional approach to signal sorting, of signal feature extraction parameters arery
manual lynseteon
the signal sorter and subsequently fine-tuned for specific
applications. This labor-intensive tuning process requires
expert field engineers and is expensive.
TRW is developing an experimental rule-based system to
facilitate the manual tuning process. Mr. Kelly observes
that the rule-based tuner has been in operation at a field
site since 1983; it is successful, and is estimated to have
saved two man-years of engineer labor to date. Mr. Kelly
described his attitude as
tuner continues to evolve; the ruleo set' is The rule-ased
growing bfrom
seventy-seven rules in 1983, to an expected two hundred
rules by February, 1984. The system is implemented in
PASCAL.
Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
 Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
2.3.7 Speaker:
STAT
National Photographic Interpretation Ctr.
Topic: AI Applications in Image Analysis
Summary: presented a preview of a report
STAT
by the Exploitation Research and Development Committee
(EXRAND) on AI applications to classified imagery. The
report is scheduled for publication by the end of December,
1983. The objective of the report will be to provide:
o guidance for focused activities by the Intelligence
Community,
o information for senior managers,
o system implications of bringing AI into image data
systems.
The details of presentation will not be
included in the present unclassified report.
Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
STAT
 Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
2.4 INVESTMENT STRATEGIES BY INDUSTRY: Some examples
illustrating how U.S. and Japanese industry are structuring
their investments in AI.
U.S.-Industry Panel:
Dr. Charles Herzfeld, Vice President and Director
of Research and Technology, ITT Corporation
Dr. Edward C. Taylor, Director of Requirements
Analysis, TRW Defense Systems Group
Ms. Norma Abel, Digital Equipment Corporation
Dr. Carl Smith, Staff Computer Scientist, Shell
Development Company
Dr. Floyd Hollister, Senior Member, Technical
Staff, Computer Science Laboratory, Central
Research Laboratories, Texas Instruments, Inc.
Dr. Carl Love, Senior Consultant, Corporate
Planning, Westinghouse Electric Corporation
Dr. Dan Schutzer, Vice President, Citibank
Topic: Why is Private Industry Investing in AI?
2.4.1.1 ITT Summary: Dr. Herzfeld commented that
twenty years ago, he may have been the first person to
mention the new science of AI in Congress . . . today, that
child.begins to walk. Industry invests in AI for three
reasons: it is probably important, useful, and profitable.
Expert systems have enormous potential to enable high
quality output from less skilled individuals. ITT is
developing rule-based systems to pinpoint flaws in the
switching systems the company manufactures. Pattern
processing is an important area that's not quite ready for
development yet. AI will be enormously important to
industry in system design, and in the programming
environment. Operational planning in business . . . first
simulating a corporation, then running it as simulated . . .
may only be ten years away. And finally, Dr. Herzfeld
observed, "indicators and warnings" tools should be
translated from the national security application into
industrial management tools for running big companies.
2.4.1.2 TRW Summary: In the view of Dr. Taylor, the
most significant factors limiting the use of advanced weapon
systems are the cognitive limitations of the human brain.
"Augustine's Law", so called, describes the exponential
distribution of skills and aptitudes. This law is
illustrated in the following graph:
Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
 Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
Percent
Successes
Achieved
Per Decile
o Kills by Fighter Pilots
o Touchdowns by NFL Backs
o Patents by Engineers
o Arrests by Policemen
o Publications by Scientists
(Adapted from N.R. Augustine
Defense Systems Management
Review, Spring 1979)
2 3 4 5 6 7 8 9 10
Statistics show that thirty percent of the total number of
successes in many competitive events are scored by the top
ten percent of competitors; the lowest three deciles
together account for only ten percent of total successes.
In combat, for example, this indicates that only the top ten
percent of operators will use advanced information
processing and display tools successfully.
It is therefore important to develop expert systems and
computer-based intelligent advisors that will emulate those
qualities that make the top decile of experts unique.
2.4.1.3 DEC_ Summary: In Ms. Abel's view, the
positive effects of expert systems on DEC have been enormous
. in her words, comparable to the effects of the
assembly line on Ford Motor Company.
Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
 Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
XCON, DEC's expert system for computer configuration, is
only one member of a large family of expert systems
currently under development at DEC . . . sales systems,
manufacturing systems, service systems are also under
development. In less than 5 years, DEC hopes to have these
systems interconnected, as illustrated below.
The Knowledge Network DEC Envisions
Before XCON, all DIGITAL's systems went through a process of
final assembly and test during which trained technical
editors configured VAX systems according to customer order
forms, out of component parts. XCON has enabled the company
to save money over the last two years by reducing the need
for this expertise. This has changed the way the
corporation does business, and DEC is introducing other
rule-based expert systems and planning systems in
manufacturing and sales. The rule-based systems are
generally written in OPS-5; the planning systems, generally
the farthest from completion, are developed in CMU's FRL, or
frame representation language, used to simulate distribution
networks.
Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
 Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
What has DEC learned about industral applications of expert
systems? Ms. Abel observed that:
o They're state-of-the-art, and therefore risky.
o They're very large software systems and need all the
support that goes with maintaining gigantic systems.
o Technology transfer is a constant, continuing process.
o Human and motivational issues of getting systems
accepted in the working environment are as difficult
as the technical issues.
o It takes a lot of corporate investment because the
systems change the way a corporation functions. With
XCON, DEC was very lucky!
o The process needs a variety of skills . . . a few AI
researchers, and a much larger number of the
applications experts already available to the
corporation.
2.4.1.4 Shell Summary: Dr. Smith from Shell
.Development Company observed that Shell evaluates emerging
technologies in terms of three basic issues:
o The Scope of the technology . . . what does it do
that is not better done with other technologies?
o The Prospectus . . . what do we expect to gain from the
technology?
o How much will it Cost to achieve the technology?
Shell considered three major components of AI for internal
applications: Robotics, Natural Language, and Expert
Systems. Dr. Smith indicated that there are three things a
company can do with respect to emerging technologies:
o nothing!
o low-level surveillance, acquiring useful systems when
someone else has developed them,
o do some development itself.
In Natural Language and Robotics, Shell is taking the second
option. In Expert Systems, Shell is taking some initiative,.
and has developed two expert systems . . . one of which is
called NDS, a communication network fault diagnostic
system developed for Shell by a contractor and still in the
prototype stage.
Dr. Smith asserts that he is enthusiastic about AI,
particularly about the tools AI has produced: the graphics
environments of the Lisp machines and the Methodology of
Expert Systems as a way of organizing heuristic procedures.
Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
 Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
2.4.1.5 TI Summary: Texas Instruments, Dr. Hollister
observed, is probably the world's third largest computer
company. TI manufactures its own chips; collects, processes
and interprets oil exploration data; manufactures metallized
and electronic products as well as consumer products like
Speak-and-Spell. TI views AI as a technology as
revolutionary as the transistor and the company invests
several million dollars in AI research each year. Since
1978, Natural Language has been, a major focus of TI's AI
research. TI's overall objective here is to enable people
to interface with computers, data bases, and other
information-processing products without requiring
specialists in data base programming to be on hand. Other
AI activity areas at TI are Expert Systems, Speech
Understanding Systems, Planning Systems and Symbolic
Computing.
TI and Western Digital are probably the major providers of
data processing services to the oil industry. TI has
experts who interpret "G-LOG data" to infer geological
structure, and the company would like to capture that
expertise in computer programs. As another potential
application area, TI manufactures its own silicon for
computer chips. There are a few TI employees in
"Augustine's upper decile" when it comes to adjusting the
process to produce pure silicon; that valuable expertise
should also be captured. Automated chip design for
testability, and automated chip inspection and flaw
detection are also expert systems targets at TI.
TI relies on robotics for production cost reduction in the
process of assembling and testing calculators.
In summary, Dr. Hollister observed, TI justifies its
investment in AI in terms of expected improvements in
existing products as well as innovative new products. The
impact from AI is expected to be felt in the late 1980's and
1990's.
2.4.1.6 Westinghouse Summary': Dr. Carl Love commented
that Westinghouse's investment in AI is almost exclusively
in developing knowledge-based systems technology. For
scientific research, Westinghouse is relying on the
university community to do the job.
Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
 Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
Today, Dr. Love observed, the US is a knowledge-based
economy. The successful manufacturing operations aren't the
largest, but the smartest . those which use knowledge.
Westinghouse has a strategy with respect to AI. Its general
principles are:
o The company can': sit back and wait while AI is
developed by others, then go out, buy the technology,
and drop it like a bomb on the company's problems.
Realistically, the company has to learn the technology
by doing it.
o Westinghouse should develop links with leading
universities; the company has derived true benefits
from its relationship with CMU.
o A company has no alternative but to provide training
for its own people. Westinghouse has a substantial
training program in place and is looking at a
component for training program managers, in addition to
training the people who will actually do the work.
o Westinghouse is establishing two or three internal AI
Centers of Excellence, as well as a distributed base of
technical competence in AI.
Generic Westinghouse expert systems application projects
include:
o order entry systems - Westinghouse is learning a lot
by talking with DIGITAL, about treating situations
where complex combinatorics are required and where
"humans find infinitely many ways to screw things up."
o the areas of equipment testing, equipment diagnostics,
and documentation of software.
o expert knowledge capture, as for example in the case of
experienced power systems engineers nearing retirement
age.
o computer perception, as for example in automated
inspection of welds.
o manufacturing and scheduling, a joint project with CMU.
Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
 Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
As to marketable AI products, Westinghouse sees enormous
opportunities for embodying technical service products in
expert systems which, unlike consultants, are readily
portable; also, in systems which support human operators in
controlling nuclear facilities. In addition, Dr. Love
believes industrial knowledge-based systems may be
instrumental in achieving substantial reduction in
production costs.
2.4.1.7 Citibank Summary: Dr. Dan Schutzer described
Citibank as a business characterized by:
o rapidly changing environments,
o need for timely information,
o requirements for data capture and decision support,
o large volumes of data to analyze,
o scarcity of experts.
In bond transactions, for example, the Citibank analyst
searches rapidly through large volumes of data for
profitable opportunities. In the Citibank working
environment, Dr. Schutzer believes AI technology will make
contributions in the areas of:
o capturing expertise in making profitable bond trades,
o capturing expertise in making long-term predictions,
o man-machine interfaces,
o development environments for rapid prototyping of
experimental systems.
2.4.2 The Japanese Fifth Generation Computing Project
Speaker: Dr. John Alan Robinson
Professor of Computer and Information
Science, Syracuse University
Subject: Japan's Fifth-Generation Computing Project:
Objectives, Status, and Prospects
In describing the Japanese Fifth Generation project, Dr.
Robinson drew the following analogy: Isaiah Berlin
has contrasted the Hedgehog and the Fox . . . the Hedgehog
knows one big thing, the Fox knows lots and lots of little
things. The central thesis of the Japanese Fifth Generation
Project asserts that lots and lots of the little pieces that
constitute modern computer programming and software
engineering can be unified by the one big idea of Logic
Programming.
Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
 Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
In Dr. Robinson's view, the Japanese project is a "symbolic
Sputnik"; the U.S. could be doing it too, but we haven't
accepted the idea that logic ought really to be at the
center of computer science.
The Institute for New Generation Computer Technologies
(ICOT) is the central facility for the Japanese project
which was initiated in 1982 and directed by Dr. K. Fuchi and
Dr. Moto-Oka. The staff consists of between forty and fifty
very capable and highly motivated computer scientists. The
project has borrowed heavily from the European developments
in logic programming and PROLOG, notably from the work of
Robert Kowalski and David Warren.
Ambitious and very detailed project plans have been
publicized by ICOT. The goal of the project is to develop
by the 1.990's a super-computer with 1012 bytes of main
memory and capable of 109 logical inferences per second
(lips)... The machine language may combine both logic and
Lisp. The project, has seven major themes and more than two
dozen sub-themes, shown in the following table.
basic application systems
1-1)
Machine translation system
1-2)
Question answering system
1-3)
Applied speech understanding system
1-4)
Applied picture and image understanding system
1-5)
Applied problem solving system
Basic software systems
2-1)
Knowledge base management system
2-2)
Problem solving and inference system
2-3)
Intelligent interface system
New advanced arch:._ctur-,
3-I)
Logic programming machine
3-2)
Functional machine
3-3)
Relational algebra machine
3
3-4)
Abstract data type support machine
3-5)
Data flow machine
3-6)
Innovative von Neumann machine
Distributed function architecture
4-1)
Distributed function architecture
4-2)
Network architecture
4-3)
Data base machine
4-4)
Highspeed numerical computation machine
4-5)
High-level man-machine communication system
VLSI technology C
S-1)
VLSI architecture
7
5-2)
Intelligent VLSI CAD system
Systematization technology
6-1)
Intelligent programming system
6-2)
6-3)
Knowledge base design system
Systematization technology for computer archi-
tecture
6-4)
Data base and distributed dzta base system
Development supporting
7
7-I)
Development support system
"chnology
Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
 Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
Many of the sub-themes are major Al projects in their own
right, such as speech and picture understanding systems.
ICOT has developed each of these project plans in
considerable detail.
In Dr. Robinson's view, this complex roadmap camouflages the
true simplicity of the overall plan. During the first phase
(1982-1985), Dr. Fuchi intends to develop a PROLOG
workstation. This will be ICOT's version of the U.S. Lisp
Machine, and will be an essential tool for subsequent phases
of the project. During phase two (1985-1989), they intend
to build a parallel inference machine. During the third
phase (1989-1992) they intend to develop the "super-parallel
inference machine," described above.
In summary, Dr. Fuchi has placed his bet on the central idea
of Logic Programming. If they continue on this path, Dr.
Robinson believes they will get where they intend, at least
in terms of their engineering goals. With the ambitiously
stated AI goals such as speech and picture understanding,
success is less predictable . . . here they'll encounter the
same sorts of difficulties that we do. Dr. Fuchi's adoption
of Logic Programming is to be commended, in Dr. Robinson's
view. Many of the problems of proving correctness of
programs simply go away, with this approach. If the
specifications are formulated in logic and the programs are
deduced from the specifications, then the programs are bound
to be as correct . . . or as faulty . . as the original
specifications. Logic Programming is also ideal for
extending the concepts of relational data bases, for
knowledge representation in expert systems, and for natural
language parsing. In Dr. Robinson's opinion, the Europeans,
and now the Japanese, have been quicker to grasp the
advantages of logic programming for such applications than
the U.S. has been.
Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
 Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
2.5 INVESTMENT STRATEGIES BY DEFENSE: A summary of DARPA's
Strategic Computing Program.
Speaker: Dr. Robert Kahn
Director, Information Processing
Techniques Office, DARPA
Topic: DARPA Strategic Computing Program
Sum ma y: Dr. Kahn observed that the DARPA strategic
computing project is aimed at Defense needs for intelligent
machines. The major program goals are:
o to speed up current logic machines by four orders of
magnitude, from the present capability of 104 to a
target capacity of 108 logical inferences per second.
o to develop generic software packages for computer
vision, speech, intelligent data bases and other
functional areas.
o to focus technology on three military demonstration
systems . . . naval battle management, autonomous land
vehicles, and automated aircraft cockpits.
o to increase the number of qualified faculty and
students in computer science.
The U.S. Government is concerned with supercomputers for
symbolic computing and for numeric computing as well. There
are supercomputing programs in each of these areas.
The DARPA program is concerned mainly with advancing the
state-of-the art in AI and symbolic computing rather than
with numerical computing. A Defense Science Board task
force, headed by Professor Joshua Lederberg, is preparing
reports on:
o high impact areas of symbolic computing within Defense,
o requirements on the technology base, imposed by the
goals of the supercomputing project,
o how to introduce AI and supercomputing into Defense,
o what the Defense investment strategy should be, in
regard to supercomputing.
These reports should be ready in 1985.
On the issue of numerical supercomputing, the White House,
represented by Dr. Keyworth of OSTP, has set up a Federal
Co-ordinating Committee on Science and Engineering
Technology Policy, known as the Picksett Committee, which is
addressing and reporting to OSTP on these issues.
Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
 Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
In planning the strategic computing program, DARPA developed
a requirements analysis by matrixing six major military
applications against twelve generic AI/Symbolic Computing
functionalities. This matrix is shown below.
MATRIX OF SYSTE=M CAPABILITIES vs.
MILITARY APPLICATIONS
AUTONOMOUS
VEHICLE
6ATTLE
MANAGEMENT
& ASSESSMENT
PILOT'S
ASSISTANT
TERMINAL
HOMING
AUTOMATED
DESIGN &
ANALYSIS
WAR
GAMING
VISION
R
0
R
SPEECH
0
R
NATURAL LANGUAGE
R
R
0
R
INFORMATION FUSION
R
R
R
'
PLANNING ? REASONING
R
R
R
R
R
SIGNAL WTREPREETAUON
R
R
R
NAVIGATION
R
R
SIMULATION/MODELING
R
R
R
R
GRAPHICS/DISPLAY
R
R
R
R
DB/IM/KS
R
R
R
R
R
R
DIET. COMMUNICATION
R
R
R
R
SYSTEM CONTROL
R
0
R
R
Battle Management emerged as the "richest" problem area,
requiring essentially all the AI functionalities;
Intelligent Data Handling referred to as Data Bases and
Information Management and Knowledge Systems (DB/IM/KS),
emerged as a functionality required by all the military
applications.
DARPA's strategy will be to stimulate technology growth by
making computing tools readily available to the academic and
industrial research communities, and by allowing these
communities to procure the tools they need. In the
microelectronics area, DARPA intends to emphasize a "fast
turnaround program," within which chip design and
multiprocessor designs can be submitted to fabrication
facilities over the ARPANET, (the Mead-Conway approach),
with production and delivery to the designer targeted for
four to six weeks.
Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
 Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
Modularity of AI systems is another goal; DARPA would like
to make it fairly simple to "plug together" the vision sub-
systems, speech sub-systems, and knowledge based sub-systems
emerging from AI research into integrated AI systems.
Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
 Approved For Release 2008/09/04: CIA-RDP86M00886R000500040010-5
2.6 THE TACTICS OF AI TECHNOLOGY TRANSFER: The pragmatics
of starting and maintaining AI R&D facilities.
Panelists:
Mr. George Lukes, Physical Scientist, Research
Institute
U.S. Army Engineer Topographic Laboratories
Dr. James S. Albus, Chief, Industrial Systems
Division National Bureau of Standards
Image Research Scientist STAT
Office of Research and Development, Central
Intelligence Agency
Dr. Jude E. Franklin, Manager, Navy Center for
Applied Research in Artificial Intelligence
Dr. Northrup Fowler, III, Computer Scientist,
Rome Air Development Center
Dr. David Brown, Assistant Director, Advanced
C,-)mputer Systems Department, SRI International
Topic: Wh: . are the P_~.,blems and Requirements for
St