-317-
Artificial Intelligence: A Tool for System Dynamics
Janet Marjorie Gould
Massachusetts Institute of Technology
ABSTRACT
This paper presents the findings of my research in artificial intelligence
applications for system dynamics. The sudden appearance of microcomputers in
homes, schools, and businesses has opened an opportunity for dissemination of
system dynamics to a wider audience than we could ever hope to reach with the
earlier computer technologies. This opportunity should not be lost by
clinging to obsolete, or soon to be obsolete, technologies. User-friendly
micro-based software should be immediately available for those individuals,
schools, and corporations who are interested in systems thinking. The demand
for such systems far surpasses the current supply. Artificial intelligence
software is now available for microcomputers. This new software development
can significantly improve current and future systems for the novice and the
experienced system dynamicist.
INTRODUCTION
Artificial intelligence has recently become one of the most publicized fields
of computer technology. This success has many implications for system
dynamicists. There are many similarities between current system dynamics and
artificial intelligence research. The problems being solved by artificial
intelligence for complex physical systems are very similiar to problems system
dynamicists have been solving for complex social systems. This paper presents
an introduction to basic artificial intelligence research and possible
applications for system dynamics. A more detailed study will follow ina
later paper.
The recent commercialization of artificial intelligence, especially expert
systems, will eventually allow some of the simpler ideas of system dynamics to
be disseminated to a wider audience. The lack of an available human expert
will be compensated for by an available artificial assistant.
I will discuss some of the current research in artificial intelligence and
particularly that research which will be potentially useful for system
dynamics. I will begin with a brief discussion of the definition of
artificial intelligence. A brief description of current artificial
intelligence topics is supplied with primary emphasis given to expert sytems.
-318-
Artificial intelligence may give us the opportunity to closely examine the way
in which we present system dynamics to others. It may help us to develop a
more innovative approach. Explanation and discussion of these new
technologies will yield important advances in our research.
WHAT IS ARTIFICIAL INTELLIGENCE?
The leaders of the field of artificial intelligence have offered different
definitions of what it is and should be. A widely agreed upon definition is
not easily found. Below are listed some rough definitions of intelligence
provided by major researchers in the field.
A distinetive aspect of what we call intelligence is
the ability to solve a wide range of new, different kinds
of problems. (Minsky 1985, p. 128)
Artificial intelligence is part of the grand attempt to
understand thinking. The programs we write are
experiments...the real results will be a new kind of
understanding of ourselves, an understanding that is
ultimately much more valuable than any program. Being able
to learn from experience and apply that knowledge in
relevant situations is an important step toward actual
intelligence. (Schank 1985, p. 155)
Artificial intelligence is concerned with extending the
capacity of machines to perform functions that would be
considered intelligent if performed by people. (Papert
1980, p. 157)
Artificial intelligence isn't about creating smart
computers. It is about something much more interesting:
intelligence. Mind. The nature of thought. -- John Seely
Brown, Xerox Corporation (Waldrop 1985, p. 39)
Artificial intelligence is the study of ideas that enable
computers to be intelligent. (Winston 1984, p. 1)
One underlying debate concerns how broadly to define artificial intelligence.
Roger Schank, from Yale University, seems to have one of the stronger opinions
about what artificial intelligence should be. Schank believes that artificial
intelligence focuses on intelligence, something that remains mysterious and
elusive. Most good artificial intelligence programs aren't terribly useful,
and many very useful "smart" programs aren't artificial intelligence at all.
He thinks that rule-based programs (a type of expert system) "do not attempt
to reason the way a human expert would and therefore are not a true
application of artificial intelligence" (Schank 1985, p. 152). Other
researchers have equally strong opinions, but as a rule seem to take a much
broader definition of what type of research is included in the field of
artificial intelligence.
For the purpose of this paper, I will take the broader view of artificial
intelligence, one which would include rule-based, expert systems.
-319-
CURRENT ARTIFICIAL INTELLIGENCE RESEARCH
Figure 1 shows many of the areas of research which have been labeled
artificial intelligence. As mentioned earlier, whether or not they really are
artificial intelligence is debatable, but for now we will except them as basic
research moving in the direction of machine reasoning.
. VISION
NATURAL
LANGUAGE
COGNITIVE —sOARTIFICIAL
UNDERSTANDING INTELLIGENGE
/ N _
ROBOTICS SYSTEMS A- 3352
Figure 1 Artificial Intelligence Research Areas
Expert systems, natural language understanding, and robotics, are the major
areas of artificial intelligence which are being commercialized. There has
been a concerted effort to make artificial intelligence available to the
general public. This is being accomplished by creative software development. i
Simplified presentations of artifical intelligence are attracting considerable |
public attention. These ideas should not go unnoticed. Simplified
presentations of key system dynamics concepts could help us disseminate our
ideas to the general public.
Robotics is being successfully used in large manufacturing plants around the
world. There are robots that walk and adjust their height for low doorways,
robots that are skilled assembly line workers, and robots that walk up stairs.
I find robotics fascinating, but not clearly relevant to the field of system
dynamics at this time.
Vision is another active area of artificial intelligence research. Vision is
finding wide application in the military. How is a threatening military tank
different from a native produce truck or even the surrounding terrain? These
types of questions are vital for understanding object differentiation on the
human as well as machine level.
-320-
The study of cognitive science or understanding is particulary interesting
especially for those who are concerned with educational uses of system
dynamics. Studying how a computer might learn will also allow us to gain
insight into how humans learn. Human learning can also give us insight into
how computers might be programmed to learn. Cognitive understanding explores
our use of language and our memory from a semantic and syntactic perspective.
It is quite clear that our brains allow us to easily differentiate between
junk, a pile of useless things or a Chinese ship, based on the context in
‘which they are discussed. Giving a computer the ability to differentiate in
this way is an extremely interesting and complex challenge. The human memory
is extremely powerful tool for complex learning and creating of new ideas.
The way in which system dynamics is taught at the university level could
certainly be enhanced by a better understanding of the way students learn.
And anyone who is interested in introducing system dynamics at the elementary
or high school level, would certainly benefit by exploring research in
cognitive science.
Natural language is becoming commercially available on small enough systems
that the general public can begin to participate with applications of their
own. Natural language research includes developing interfaces with the
computer that will allow a user to communicate with the computer almost
entirely in the English language. These interfaces can be utilized by
entering information through the traditional keyboard or verbally. There are
some voice recognition systems available, but currently these are severely
limited by not recognizing more than one person's voice. The goal of future
research would be to allow any user to communicate with the computer. Natural
language communication through the keyboard is commercially available and will
be one of the most popular artificial intelligence products of the near
future.
Expert Systems are the latest commercially successful artificial intelligence
product. Expert systems try to emulate the knowledge of an expert. Some of
the well known systems are MYCIN, PROSPECTOR, and XSEL. These systems are used
to identify bacterial infections, for geological surveying, and to configure
computer systems, respectively. MYCIN has been useful for physicians making
complex medical diagnoses, PROSPECTOR for finding specfic occurances of
certains ores, and XSEL for the salesperson faced with the complexities of
designating the correct components and peripherals for a Digital Equipment
Corp. computer system. XSEL is based on an earlier system XCON which is used
to configure computers like the VAX 11/780 or PDP-11.
A small subset of expert systems research is being done in rule-based systems.
Other systems are based on nets and frames, which are more complex
representations than rules. Frames are similar to records in Pascal. That is
an object, such as an address, is associated with a person's name, street,
city, state, and zip code. A net can connect different frames together.
Addresses of different computer stores would be connected to restaurant
addresses for a directory, such as the Yellow Pages, to forma net.
~-321-
Rule-based systems refer to the way in which knowledge is represented in these
particular types of expert systems. They are based entirely upon rules of the
if-then type. For example, if it is raining, I will bring my umbrella. The
decision tree for this type of statement would begin with the statement, it is
raining. A response of either true (it is raining) or false (it is not
raining) would be determined. True and false would be separate branches of
the tree. If true, then I will bring my umbrella, and the tree would branch
out from there. If false, the other branch would be chosen and continue from
there.
M.1
Teknowledge, Inc.
Copyright (c) 1984, All rights reserved.
Version 1.0, Serial Number 0165
M.1> load wine.
M.1> go.
Is the main component of the meal meat, fish or poultry?
=> meat.
Does the meal have veal in it?
DONO.
Does the meal have a sauce on it?
>> yes.
Is the sauce for the meal spicy, sweet, cream or tomato?
=> why.
Your answer to this question will help determine whether the
following rule is applicable in this consultation.
ruled:
if not main-component=fish and
has-sauce=yes and
sauce=tomato
then best-color=red.
The following knowledge base entries are also being used:
rule2s (a rule)
rule4S (a rule)
s the sauce for the meal spicy, sweet, cream or tomato?
=> show ruleZé.
rule2é:
if best-color=red
then recommended-color=red.
the sauce for the meal spicy, sweet, cream or tomato?
unknown.
you generally prefer red or white wines?
Reprinted with permission of Teknowledge, Inc.
A- 3348
Figure 2 Wine Program ~- the use of "why" and "unknown"
~322-
Figures 2, 3, and 4 are examples of the printed ouput of a rule-based system
ealled M.1 that contains knowledge about the correct wine to be served with a
meal. I chose this example so it would be easy to follow. M.1 is available
from Teknowledge, Inc. It includes some examples of expert systems such as
the wine program, a banking program, an engineering program, etc. M.1 is
designed with two separate components, the inference engine and the knowledge
base. The inference engine manipulates the knowledge supplied by the user and
provides a simplified parser to be able to understand English responses. The
knowledge is created by the user with a series of if-then rules. M.1 will
accept about 300 rules.
M.1's inference engine is backward chaining. It begins with a conclusion and
searches backward trying to satisfy specific conditions. For example, in an
animal identifying program, suppose the conclusion is that the animal is an
elephant. What conditions must be satisfied to make this true? If the animal
has a trunk, and a tail, and is a mammal, then it is an elephant. Tracing
backward from the conclusion that the animal is an elephant, we must examine
the conditions that must be satisfied about the trunk, tail, and mammal. For
example, if the animal nourishes its young with its own milk and has hair,
then it is a mammal. We then need to trace back further to the conditions
about hair and nourishment. This process continues until all conditions are
satisfied to conclude that the animal is an elephant.
white
the flavor of the meal delicate, average or strong?
* why.
Your answer to this question will help determine whether the
following rule is applicable in this consultation.
rules
if tastiness=delicate
then best-body=light cf B80.
The following knowledge base entries are also being used:
rulel4 (a rule)
rule4s (a rule)
Is the flavor of the meal delicate, average or strong?
Ho average.
Do you generally prefer light, medium or full bodied wines?
>» medium.
you generally prefer dry, medium or sweet wines?
medium.
wine zinfandel (90%) because rules4
wine cabernet-sauvignon (90%) because rules4
wine pinot-noir (64%) because ruless
wine gamay (64%) because rule4S
wine burgundy (20%) because rulesé
wine = valpolicella (30%) because ruleSi
Reprinted with permission of Teknowledge , Inc.
4-334,
Figure 3 Wine Program - rules and results
-323-
M.1 is essentially ignorant until the user provides the rules. Figure 2 shows
an example of an if-then rule, "rule9". The wine program essentially is a
dialogue with the system. After the "go." statement the program begins a
series of questions which will help it traverse the tree of rules to come to a
final conelusion about which wine is appropriate. The user is allowed to
question the system at any time by typing "why.". In response, the system
will display the rule it is currently using to try to reach a conclusion. The
user may ask the system to show what is contained in a particular rule, such
as "rule26" in Figure 3. If the answer to one of the program's questions is
not known, the user may type "unknown." and the system will ask another
question.
M.i> show.
wine = zinfandel (90%) because ruleS4
wine cabernet-sauvignon (90%) because rules4
wine pinot-noir (64%) because ruless
wine gamay (64%) because rule4s
wine burgundy (30%) because rulesé
wine = valpolicella (30%) because ruleS3
recommended-color = red (90%) because rulezé
best-color = red (90%) because ruled
main-component = meat (190%) because you said so
has-veal = no (100%) because you said so
has-sauce = yes (100%) because you said so
sauce was sought, but no value was concluded.
preferred-color = white (100%) because you said so
recommended-body = medium (64%) because rulelt8 and rulets
recommended-body =
wine zinfandel (90%) because ruleS4
wine cabernet-sauvignon (90%) because rules4
wine pinot-noir (64%) because ruleSs
wine gamay (64%) because rule4s
wine burgundy ) because rulesé
wine valpolicella (30%) because ruleS3
recommended-color = red (90%) because rule2é
best-color = red (90%) because ruled
main-component = meat (100%) because you said sa
has-veal = no (100%) because you said so
has-sauce = yes (100%) because you said so
sauce was sought, but no value was concluded.
preferred-color = white (100%) because you said so
recommended-bady = medium (64%) because rulei8 and rule1S
recommended-body = full (20%) because ruleléd
recommended-body = light (30%) because rulel4
ii bu we
best-body = medium (60%) because rule3
best-body = full (30%) because rules
best-body = light (30%) because rules
tastiness = average (100%) because you said so
preferred-body = medium (100%) because you said so
recommended-sweetness = medium (100%) because rule4l
best-sweetness was sought, but no value was concluded.
preferred-sweetness = medium (100%) because you said so
M.1> Reprinted with permission of Teknowledge, Inc.
Figure 4 Wine Program - explanation A- 3350
-324-
In Figure 3, "rule2" is displayed. At the end of the rule is printed, "cf
80". This indicates that the certainty factor is 80. The certainty factor,
supplied by the person who provided the knowledge, indicates that if the
response, "delicate", is received when the user is questioned about tastiness,
then it is about 80% likely that the best body of the wine is light. The last
question in Figure 3 has the response, "medium". The last statement of figure
4 shows that the preferred sweetness of the wine is definitely medium (100%),
"because you said so". That is, I supplied the information with my response
of "medium". I did not provide a certainty factor, so the system assumes a
factor of 100. The last six lines of Figure 3 show which winés the program
would recommend. Figure 4 shows the user what conclusions were reached along
the way and which rules they were based on. Those rules could then be
examined for further understanding.
What does all this information imply for system dynamics? The goal is to
harness the power of the computer to manipulate heuristics as well as specific
algorithms and numeric data. The ability to query a system about its line of
reasoning and to be able to use the answer "unknown" in response to a question
are extremely powerful ideas. We all realize that in conventional programming
methods this power is not available. During a standard interactive program a
response of "unknown" would (unless specifically programmed as an acceptable
response) cause an error in the program. And certainly attempting to ask
"why" or how the program was processing information would be futile. These
capabilities present us with the first attempts to model human reasoning.
Humans can explain their paths of reasoning. We are capable of reaching
conclusions although certain information is unknown.
ARTIFICIAL INTELLIGENCE AND SYSTEM DYNAMICS
In order to make artificial intelligence available to the general public, an
effort has been made to make certain ideas and software packages available
which one can understand without a Ph.D. in the field. The field of system
dynamics has faced some difficulty in being accepted by other academics and
certainly by the more general audience because of perceived complexity. As in
other sciences and related disciplines, we should devote some time to
simplifying the key ideas of systems dynamics. "The central task of a natural
science is to make the wonderful commonplace: to show that complexity,
correctly viewed, is only a mask for simplicity; to find pattern hidden in
apparent chaos" (Simon 1969, p. 1). The social sciences should be challenged
to simplify certain complexities. This is what is being accomplished by
Papert through his development of LOGO microworlds to educate young students
about math, physics, etc. System dynamics should begin to take this approach
not only for students, but for corporate executives and anyone else who
displays an interest in our research. We should be ready to take advantage of
new artificial intelligence applications as they emerge over the next few
years. As an academic discipline and a public policymaking perspective,
system dynamics could thrive on these developments. "Within the next two
decades, new capabliities such as computer-driven video disks and video
tapes--along with clever use of computer instructions games and self-paced
instruction packages should push the field to new levels of accessibility and
appeal" (Gould 1983, p. 19).
My current research is exploring combinations of natural language and expert
systems for system dynamics. The software currently available for system
=325-=.
dynamics models has the same drawback as most other software, it does not
allow standard English input. By designing interfaces that incorporate a
“system dynamics parser" with expert knowledge and the ability to generate
dynamic simulations, we would then have a system usable by people with
different levels of expertise.
One of the major difficulties in simplifying the system dynamics approach is
the in-depth understanding one needs to become a competent model builder and
interpreter of model behavior. These abilities are not gained after one
course at the graduate level. It can take many years of model building before
one is a good modeler. Some people never become competent model builders even
after years of practice. The skills of a Jay Forrester are rarely mastered.
If Forrester's knowledge were available in a database, we could eventually
develop a computer system which would generate its own outstanding models. In
artificial intelligence, knowledge engineers try to extract knowledge from
experts to be entered into an expert system. The MYCIN program, mentioned
earlier, has the knowledge of more than one physician. This extensive
knowledge can then be used to guide those who do not know "all the right
answers", Of course, one still needs the right questions. Although MYCIN
contains information about infectious diseases, the user would not be able to
interact with the system without knowing the correct terminology.
Knowledge acquisition tools are being developed, but much more slowly than
expert systems themselves. It has become clear that no knowledge base is ever
complete. A dynamic process is needed for changing the set of rules. As new
information is needed, the knowledge base must be changed. Unfortunately, the
knowledge engineer is often unavailable, so the user is expected to improve
the system. Knowledge acquisition systems are needed. It would seem
reasonable to expect that future expert systems will be supplied with
knowledge acquisition tools. TEIRESIAS was developed to allow a user to
update and expand the knowledge base of MYCIN. A system is being developed to
improve the current version of XCON. As these systems are developed, it will
certainly be useful for system dynamacists to be aware of them and be ready to
create their own systems.
Papers by Davis and Kuipers of MIT on structure and behavior of physical
systems are interesting for our work in social systems. Kuipers paper is
"concerned with the qualitative simulations of physical systems whose
descriptions are stated in terms of continuously varying parameters. The
examples presented demonstrate a representation for qualitative reasoning
about causality in physical mechanisms. The system as described in this paper
has been completely implemented in MACLISP. The structural description is
essentially a qualitative form of a differential equation, specifying a set of
parameters which characterize the state of the mechanism and a set of
constraints holding among the parameters. Qualitative simulation produces a
behavioral description which specifies the ordinal relationships and
directions of change of the parameter values at each point in time" (Kuipers
1984, pp. 189).
Davis is interested in troubleshooting electronic circuits. Davis' research
"goal is to develop a theory of reasoning that exploits knowledge of structure
and behavior. The initial focus is troubleshooting digital electronic
hardware, where we have implemented a system based on a number of new ideas
and tools. We have developed languages that distinguish carefully between
-326-
structure and behavior, and that provide multiple descriptions of structure,
organizing it both functionally and physically. We have argued that the
concept of paths of causal interaction is a primary component of the knowledge
needed to do reasoning from structure and behavior. We deseribe a new
technique called constraint suspension, capable of determining which
components can be responsible for an observed set of symptoms. We show that
the categories of failure, previously derived informally, can be given a
systematic foundation. We show that the categories can be generated by
examining the assumptions underlying our representation" (Davis 1984b, p. 3).
Davis has also developed a simple circuit drawing system that permits
interactive entry of pictures which represent the circuit structure and are
then translated in to the proper Lisp code. Davis is creating a type of
expert system for physical systems which could eventually be very useful for
social systems. Of course, at this level physical systems are easier to
manipulate than social systems because they are more easily quantified.
Davis' paper has lead me to theorize about the type of systems we can expect
to be using in the future.
Figure 5 is a representation of the type of expert system which could
eventually be available for system dynamics. This system would initially be
designed for educational purposes, using very simple system dynamics generic
structures or models. These simple models are currently a major research
focus designed to demonstrate that the same basic structures are often found
in many different environments. An interesting example of this which I
recently encountered concerns similar structures underlying the desire for new
members in a church and a health maintenance organization. At first, one
would probably not recognize the similarities between a church and a health
maintenance organization, but a very simple generic structure can be developed
that underlies both. The desire for more members puts pressure on the
ministers and doctors to interact with more people. This brings in more
members. The more members there are the less time the ministers or doctors
have to spend with them. This brings about dissatisfaction among members.
Members will quit unless more ministers or doctors are hired.
My basic theory is that a user can communicate with a system dynamics model
through an expert system. The rule-based interface would be knowledgeable
about the differences between the various generic structures to which it has
access. One could communicate with the interface either in English or
graphically (much like Davis’ system for circuits). A user would begin with a
particular problem that, when understood by an expert in system dynamics,
would be associated with a particular type of generic structure. The
interface would take information from the user and decide which generic
structure to use. The interface would then pass control on to a more complex
expert system, the simulator, which contains information about specific
generic structures. In the simplest configuration the fault detector would
have knowledge about preferred types of behavior that are associated with a
specific generic model. When that behavior is not displayed based on the
user's parameter definitions or changes in structure, a fault or error
detector would attempt to determine from the structure, where the fault lies.
Control could be passed between the simulator, fault detector, and finally to
the user for assistance.
A more sophisticated version of Figure 5 could be used to develop simple
models. The user would provide information to the fault detector about the
-327-
type of behavior that the user expects based on his model structure. After
the model is created (in English or graphics) and simulated, the fault
detector would act as a helper to the user in identifying problem areas in the
model.
Gite SSS Seas PrLat See AeoEto oS peee BS CSRS os 1
i
' i! '
t
GENERIC |}! GENERIC Generic | 1
1 | FAULT 11] FAULT
STRUCTURE} | STRUCTURE STRUCTURE] |
1 DETECTOR) |simucator|!'|OETECTOR IsimuLaTor SIMULATOR | !
Lenmeews anf pawns Mwawmeneel: wen cowed eueeaenpeenosooe 4
USER INTERFACE
Knowledge
about different
generic structures
A- 3353
Figure 5 System Dynamics Expert System
A simple example may help to illustrate the idea shown in Figure 5. Suppose a
student was interested in "playing with" a model. Suppose the student, while
interacting with the interface, created a one-level model. The student's
model would then become an object to simulate (in the generic structure
simulator). Suppose the student made an error and thought that the model
should exhibit oscillations. The model would simulate and display growth.
The fault detector is then faced with a discrepancy between the student's
anticipation and the actual behavior. The detector would also contain
specific knowledge about the behavior of system dynamics structures. The
fault detector would analyze the structure of the model in the simulator and
realize that the specific structure should not oscillate. That information
would then be passed back through the interface to the student, if desired.
The interaction between the student, the interface, the simulator, and fault
detector provides an intelligent educational environment. Anticipating
behavior based on structure is a complex task associated with physical and
social systems. The complex nature of systems often creates counterintuitive
-328-
behavior. To understand how behavior is generated from specific structures,
one must have substantial experience creating and manipulating structures.
The intelligent computer environment I have suggested creates what is
necessary to encourage the learning process.
The more complex a model becomes the more knowledge the fault detector would
need in order to assist the user. Eventually knowledge about phase and gain
relationships, stable and unstable equilibria, eigenanalysis, etc. would be
needed. For now, I would envision the early prototype of the system to be
designed for a student rather than for the professional system dynamacist.
CONCLUSION
Much more research is needed before artificial intelligence can be immediately
useful for system dynamics. The areas which should be explored are cognitive
understanding, natural language, and expert systems. The work of Davis and
Kuipers is particularly noteworthy, and should be more closely examined.
Understanding the system dynamics learning and model-building processes is
vital. Interacting with the computer in English would be helpful for
disemminating the ideas of system dynamics to others.
I do not intend to suggest that the ideas represented in Figure 5 will be
available immediately. It is generally assumed that an expert system can
easily take five man-years to develop. I suggest that we build on earlier
developments. The work of Kuipers and Davis can provide us with the needed
base for developing our own intelligent systems. A rule-based system such as
M.1 provides an opportunity for system dynamacists to experiment with and
manipulate our own knowledge base. Microcomputer software such as M1, Expert
Ease, Gold Hill Lisp, etc. put the power of knowledge systems in the hands of
the individual. The opportunity is worth exploiting.
Figure 6 suggests system dynamics, education, and artificial intelligence
combining into intelligent educational systems. We realize that "in knowledge
lies the power"; but even more important is the ability to communicate
effectively and thereby pass the power on to others. System dynamics gives us
the ability to understand complex systems. Artificial intelligence gives us
insight into our own thought structures and gives us freedom to focus on the
more challenging components of learning. Education teaches us to communicate
to others about our expert knowledge. Understanding, insight, and
communication are the keys to our power as individuals and as a society. My
goal is to give individuals the power to learn through their own discoveries.
-329-
ARTIFICIAL
LLIGENT™:
Figure 6 Intelligent Educational Systems
REFERENCES
Black, R. "Artificial Intelligence for Micros", Hardcopy, March 1985, Volume
14, Number 3, pp. 43-49,
Davis, R. “Amplifying Expertise with Expert Systems," The AI Business, M.I.T.
Press, 1984, pp. 17-40.
Davis, R. "Diagnostic Reasoning Based on Structure and Behavior", M.I.T. A.I.
Memo 739, June 1984.
Fogel, L. J., A. J. Owens and M. J. Walsh. Artificial Intelligence Through
Simulated Evolution, John Wiley and Sons, Inc., New York, 1966.
Forbus, K. D. A Study of Qualitative and Geometric Knowledge in Reasoning
about Motion, M.I.T., AI-TR-615, February 1987.
Foster, E. "Artificial Intelligence: Beyond the Buzzwords", Personal
Computing, April 1985, pp. 62-69.
~330-
Gould, J. M. and D. P. Kreutzer. "New Computer Graphics Technologies for
Representing the Structure and Behavior of System Dynamics Models", System
Dynamics Group Working Paper D-3431, December 1983.
Jones, K. "Microsoft, Others Plan Expert Systems," Mini-micro Systems, July
1984, pp. 89-91.
Kuipers, B. "Commonsense Reasoning about Causality: Deriving Behavior from
Structure," Artificial Intelligence, 24 (1984), pp. 169-203.
Michaelsen, R. H., D. Michie, and A. Boulanger. "The Technology of Expert
Systems", BYTE, April 1985, Volume 10, Number 4, pp. 303-314.
Miller, R. K. “Artificial Intelligence: A New Tool for Manufacturing",
Manufacturing Engineering, April 1985, pp. 56-62.
Minsky, M. "Communication with Alien Intelligence", BYTE, April 1985, Volume
10, Number 4, pp. 127-138.
Papert, S. Mindstorms, Basic Books, New York, 1980.
Probert, D. E. "Bridging the 'Interpretation Gap' in Business Policy
Analysis," Presented at the Fourth International Symposium on Forecasting,
July 1984,
Probert, D. E. "Towards Expert Systems for Telecommunications Policy
Analysis, Computer Communications, Butterworth and Co., Volume 6 Number 2,
April 1983, pp. 58-64.
Schank, R. and L. Hunter "The Quest to Understand Thinking", BYTE, April
1985, Volume 10, Number 4, pp. 143-158.
Shiel, B. "Power Tools for Programmers," Datamation, February 1983, pp.
131-143.
Simon, H. A. The Sciences of the Artificial, The M.I.T. Press, Cambridge,
Mass., 1969.
Smith, E. T. "When Will a Computer Be Able to Communicate Like a
6-Year-Old?", Business Week, April 1, 1985, pp. 84-85.
Thompson, B. A. and W. A. Thompson. "Inside an Expert System", BYTE, April
1985, Volume 10, Number 4, pp. 315-333.
Waldrop, M. M. "The Machinations of Thought", Science 85, March 1985, Volume
6, Number 2, pp. 38-45.
Webster, R. "Expert Systems on Microcomputers", Computers and Electronics,
March 1985, Volume 23, Number 3, pp. 69-73.
Winston, P. H. Artificial Intelligence, Addison-Wesley Publishing, Reading,
Mass., 1984.
Winston, P. H. "The Lisp Revolution", BYTE, April 1985, Volume 10, Number 4,
Pp. 209-222,
