Download as ps, pdf, or txt
Download as ps, pdf, or txt
You are on page 1of 11

Real world applications of qualitative reasoning: Introduction to

the special issue

Yumi Iwasaki
Knowledge Systems Laboratory
Department of Computer Science
Stanford University

Qualitative physics has attracted much interest from the artificial intelligence research community in the
last decade. Every year since 1986 has seen two to three technical paper sessions devoted to qualitative
reasoning and modeling of physical devices in the National Conference on Artificial Intelligence. A
successful international workshop on qualitative reasoning about physical systems has been held every year
since 1987. The emphasis in qualitative physics on problems of reasoning about physical systems where
there is incomplete information, and abstraction/aggregation of models of physical systems as well as of
analysis results makes the technology relevant to many real-world industrial problems. Many applications
of qualitative physics technology have been reported in various tasks such as diagnosis, design, tutoring,
real-time monitoring and hazard identification. The purpose of this special issue on applications of
qualitative reasoning is to showcase some of the recent applications.

Broadly speaking, the research in qualitative physics is aimed at developing representation and reasoning
techniques that will enable a program to reason about the behavior of physical systems without the kind of
precise quantitative information needed by conventional analysis techniques such as numerical simulators.
Early research in qualitative physics was aimed at capturing people's ability to reason about and interact
appropriately with common-place physical situations. However, as the capability to reason qualitatively is
essential in complex engineering problems that are far beyond commonsense and require highly specialized
knowledge, the scope of the research has quickly expanded to include engineering domains.

Experts in most engineering fields routinely reason with much less information than would be required to
do detailed quantitative analysis. There are various reasons for this: Firstly, precise quantitative data may
not be available or too costly to obtain. For example, at an early stage of design of devices, precise
numerical values of the design parameters tend not to be available. Secondly, one may only need
qualitative predictions. Observing pouring rain and the steadily rising water level of a river is sufficient to
make one take measures against possible flooding without knowing the exact water level, the rate of
change, or the time the river might flood. Finally, even when detailed quantitative analysis is to be
performed to obtain precise results, qualitative reasoning is the crucial initial step that allows one to identify
quickly potential problems that warrant more detailed analysis.

Qualitative reasoning technology offers the following advantages over conventional numerical simulation:

Coping with incomplete information: When one tries to build a simulator, it is sometimes the case that
there is not enough quantitative information available to perform accurate simulation. Qualitative modeling
and simulation technology allows one to build a simulation model and perform prediction even with
incomplete knowledge of the system to be simulated.

Imprecise but correct prediction: With qualitative simulation technology, one can perform simulation
even when precise numerical model is not available since qualitative simulation will generate all possible
courses of behavior that could follow from the given incomplete specification. In contrast, in numerical
simulation, one has to assume a specific value for each numerical parameter in order to perform simulation
even if values are not known. Though numerical simulation will produce precise predictions, validity of
the predictions depends on such assumptions. It is often more desirable to see an accurate but less precise
prediction than to see a very precise one which may be incorrect.

Easy exploration of alternatives: To see the entire range of possible behaviors with conventional
numerical simulation, a large number of simulation must be performed since each run of a precise

1
simulation produces one point in an infinite, multi-dimensional space of possible behaviors. Qualitative
simulation can produce in one run all possible behaviors that could follow from the given incomplete
specification, allowing one to see the entire range of possible courses of behavior at once. In other words,
qualitative simulation can paint a rough picture of a large space of possibilities with a broad brush to give
one a insight into the overall shape of the space quickly.

Automatic interpretation: Numerical simulation produces detailed numerical outputs that describe the
behavior over time. In order to make sense of the outputs, a person knowledgeable of the phenomena being
simulated must interpret the numbers to identify important qualitative characteristics of the behavior.
Qualitative simulation produces predictions directly in terms of salient qualitative characteristics of the
behavior which requires much less effort to interpret.

1. Qualitative inference
To get a flavor of what is meant by qualitative reasoning, consider a the following everyday scenario: You
are confronted with a water-filled pan on a lit stove. You can easily predict that the pan will warm up,
which will warm the water, the water will start to boil sometime, the pan may eventually become empty,
and so on. To make these predictions, I need not tell you the exact values of the variables involved, such as
the amount of water, the temperatures of the stove and the water, or the boiling temperature. Neither do
you need to know the exact mathematical relations among the variables. Your reasoning process (assuming
you need to reason about this simple situation at all) probably goes something like the following: Heat
flows from an object with a higher temperature to another of a lower temperature. The temperature of the
water is usually less than its boiling point, and the stove temperature is usually higher than the boiling
point. So, the temperature difference will cause heat to flow from the stove to the pan and to the water.
This will cause the water temperature to rise and, eventually, the water will start to boil, etc.

Because of the lack of precision in the available information, you would not be able to say exactly when the
water will start to boil or how long it will take it to evaporate completely. Nevertheless, this type of
imprecise prediction suffices in many situations to allow people to react appropriately. Research in
qualitative reasoning is concerned with representation of such imprecise knowledge as well as inference
techniques using such knowledge to solve problems.

One note of clarification: Imprecision here should be distinguished from uncertainty. Uncertainty as in
probabilistic or fuzzy inference refers to the degree to which one believes in a certain fact. Precision refers
to the precision of the statement of the fact itself. In the above example, a given fact may be that the initial
temperature of the water is somewhere between the freezing and boiling temperatures. This is an imprecise
statement about the temperature since it does not say exactly what the temperature is. However, it is not an
uncertain statement since the statement is 100% true (or believed to be true). An example of an uncertain
statement in this case would be "it is 90% certain that the temperature is 30°C."

Another salient feature of the type of reasoning illustrated in the above example is that the description of
the predicted behavior is given in terms of the causal relations that give rise to the behavior. The notion of
causality is an essential element of human understanding of all phenomena in the world, and it is also one
of the important features of many qualitative reasoning systems, which also distinguishes qualitative
reasoning from conventional numerical simulation or mathematical analysis. In human reasoning as well as
qualitative reasoning systems, knowledge of causal mechanisms not only enables one to predict behavior
but also to provide justification for the predicted (or observed) behavior.

Causality has been an important topic for researchers of qualitative reasoning for several reasons: A causal
explanation makes causal dependency relations among events explicit, which tells one how to manipulate
the situation to influence the outcome. For example, a causal explanation of the water heating scenario will
allow one to identify possible ways to prevent the water from boiling. A causal explanation makes explicit
the knowledge of underlying physical laws or mechanisms that bring about the causal interactions. This
knowledge allows one to make generalizations and to reason about similar situations involving the same
causal mechanisms. A causal account is also much more satisfying as an explanation of why the situation
evolves as predicted than an account that simply chronicles the sequence of events, which makes qualitative
reasoning systems potentially more suitable for teaching than conventional simulation.

2
In the next section, I provide a brief explanation of fundamental ideas in qualitative physics, including
qualitative calculus and simulation as well as a representation formalism of knowledge of the physical
world. I will also touch on the issues model formulation and abstraction, which have been studied
intensively in recent years. The purpose of this brief exposition is not to give a comprehensive tutorial or
survey of the area but to provide enough basic understanding of the common techniques and issues in
qualitative reasoning for better appreciation of the articles featured in this special track. Readers are
referred to references listed at the end of this article for more information about the field.

2. Basic qualitative reasoning technique


Qualitative calculus is a mathematical tool used in many qualitative reasoning systems in order to make
inferences about behavior from imprecise information.

2.1 Qualitative arithmetic


Qualitative arithmetic is an arithmetic of signs, or more generally, of intervals. (For the sake of brevity and
ease of exposition, I will only discuss the case where signs are used. However, one can easily generalize all
the following discussions about qualitative arithmetic to cases where intervals other than positive, zero and
negative are used as qualitative values.) In qualitative arithmetic, the variables take on the values of
positive (+), negative (-) or zero (0). When one does not know the precise value of the temperature T of the
water, but knows that it is above 0°C, one can say that the qualitative value of T is +. Given [+, 0 -] as the
entire set of possible qualitative values for each variable, one can define qualitative arithmetic operators
[+], [-], [*], [/] corresponding to the arithmetic operators of the real numbers in an obvious fashion. Table
1(a) and 1(b) provide definitions of [+] and [*] as examples.

+ 0 -

+ + + ?

y 0 + 0 -

- ? - -
Table 1(a): x [+] y

+ 0 -

+ + 0 -

y 0 0 0 0

- - 0 +
Table 1(b): x [*] y

Note that, unlike the familiar arithmetic operators of the real numbers, the qualitative arithmetic operators
may not give a unique value. In the case of qualitative plus [+], the value of x [+] y can be any one of +, 0
or - if the qualitative values of x and y are + and - or vice versa.

3
In addition to qualitative arithmetic operators, there are qualitative functional relations such as M+ and M-.

M+(y, x)

or

M-(y , x)

states that y is a monotonically increasing (or decreasing) function of x without specifying the exact
functional form.

Using qualitative arithmetic operators and relations, one can describe the behavior of a system in terms of
the qualitative relations among its variables. The definitions of the operators can, then, be used to solve a
set of qualitative equations to obtain a set of possible combinations of qualitative value assignments to the
variables.

2.2 Qualitative simulation


Qualitative values give only minimal information about the state of the system being modeled. Qualitative
arithmetic becomes more useful when some of the variables are derivatives, since their qualitative values
indicate whether the variable is increasing, decreasing or steady. Instead of describing precisely the rules
for using qualitative equations for simulation, we provide below a simple example of qualitative simulation
since the rules are fairly straightforward and based on ordinary calculus of the reals.

The example is a simple mass-spring system shown in Figure 1. A block is attached to an ideal spring
resting on a frictionless table top. The variables are as follows:
x: position of the block,
x = 0 is the equilibrium position of the block when the spring is relaxed.
x > 0 when the block is to the right of the equilibrium position.
x < 0 when it is to the left.
v: velocity of the block
a: acceleration of the block
f: force acting on the block.
m: mass of the block. A positive constant.
k: spring constant. A positive constant.

0 x

Figure 1: A mass-spring system

The following two qualitative equations describe the physical laws governing the behavior of the system.

f = m [*] a Newton's second law

f = -k [*] x Hooke's ideal spring law

Given these qualitative equations, one can produce all combinations of variable values that satisfy them.
Table 2 shows all the possible combinations. Each combination represents a distinct qualitative state the
mass-spring system can be in. For example, state s 1 is the state where the block is to the right of the

4
equilibrium position (i.e. the spring is stretched) and moving right but decelerating. State s8 represents the
instantaneous state where the block is to the left of the equilibrium position (i.e. the spring is compressed)
and its velocity is zero.

s s s s s s s s s
1 2 3 4 5 6 7 8 9

x + + + 0 0 0 - - -

v + 0 - + 0 - + 0 -

a - - - 0 0 0 + + +

f - - - 0 0 0 + + +

Table 2: Qualitative states of the mass-spring system

Furthermore, because we know that v is the derivative of x and a is the derivative of v by definition, we can
also determine the possible transitions among the states. Figure 2 shows all the possible transitions among
them schematically, where the states are arranged in a 2-dimensional plane with x and v as the horizontal
and vertical axis.

s4
s7 s1

s8 s5 s2
x

s9 s3
s6

Figure 2: Transition graph for the mass-spring system

Note that the transition graph indicates a possible oscillatory behavior, which an ideal spring will exhibit if
it starts out in any state except s5. The state s5 represents the quiescent state where nothing ever changes.

As can be seen from this simple example, qualitative calculus allows one to predict some characteristics of
behavior from imprecise information about the system. It must be noted that the prediction will be
qualitative in the sense that it will not include precise quantitative information about the values of variables
or exactly how long each qualitative state will last. In the case of the above mass-spring example, the
prediction does not say anything about the magnitude of the oscillation or the frequency. Furthermore, in

5
many situations, qualitative prediction will be ambiguous in the sense that it will produce a tree or a graph
of states where a state have multiple possible next states. Such ambiguity is inherent in qualitative calculus
since the given information is imprecise to begin with.

There are several qualitative simulation systems described in the literature. QSIM [1]and QPE [2, 3]are the
best known systems that are publicly available. Since the size of the graph of states can grow exponentially
in many problems, there have been much work on limiting the number of possible behaviors produced by
qualitative simulation by using additional information. There is a considerable body of work on techniques
for limiting branching in qualitative simulation.

2.3 Knowledge representation


As techniques for reasoning with imprecise information became fairly well established, the problem of
model formulation has emerged as an important topic. While various analysis techniques, qualitative or
quantitative, allow one to make inferences about the given physical system, one must have a model to begin
with in order to use such techniques. Much work in qualitative physics research has been devoted to study
of various modeling strategies as well as representation formalism for knowledge of the physical world
needed to formulate a model.

Though simulation is a very useful tool for evaluating design alternatives in engineering design, it is not
currently used as much as it should be because of the effort involved in formulating a model. If a system
could quickly formulate an appropriate model for analyzing the particular aspect of interest, perform
simulation, and produce an interpretation of the results in an easily understandable form, it would make it
much easier for a designer to explore the space of alternatives quickly. Consequently, the designer would
be able to make better informed decisions during the design process, improving the efficiency of the
process as well as the quality of the final design.

Main concerns dictating the design of knowledge representation formalisms in qualitative physics are
generality and modularity.
Generality: The representation formalism should be general enough to be suitable for
representing a wide variety of physical phenomena, in different domains as well as based
on different theories about the world.
Modularity: The representation should be modular. Each module should be small enough
and independent enough so that they can be reused in different combinations to compose a
complete models of different physical situations.

An important paradigm in automatic formulation of models is compositional modeling, which is based on


the idea of constructing a model by composing pieces, each of which describes a specific aspect of some
physical phenomenon. Compositional modeling is an effective method for automatically formulating a
behavior model of a physical system that can be adequately modeled as a lumped-parameter system.

In the compositional modeling approach, a system is provided with a library of model fragments. Each
model fragment is a representation of a conceptually independent physical phenomenon. Here, the word
phenomenon is used in a very general sense to include the notions of physical objects, specific
combinations of physical objects, behavior characteristics of objects or combinations thereof, and physical
processes. Each model fragment consists of the following parts:
participants: The types of objects that must exist for the phenomenon to take place
quantities: The attributes of the phenomenon represented by the model fragment.
conditions: The conditions that must be satisfied by the participants in order for the
phenomenon to actually take place.
consequences: The functional relations (which can be qualitative) that will hold among the
attributes of the participants or the phenomenon itself.

Figure 3 shows an example of model fragment representing the physical process of fluid-flow.

6
ModelFragment Fluid-flow
:subclass-of (fluid-process)
:participants
source :type liquid-container
sink :type liquid-container
path :type pipe
:quantities
flow-rate :dimension mass-rate-dimension
:conditions
(connects path source sink)
(pressure-in sink) < (pressure-in source)
(amount-of-liquid-in source) > 0
:consequences
(M+ flow-rate, (Pressure-in source) - (Pressure-in sink))
(I- (amount-of-liquid-in source) flow-rate) 1
(I+ (amount-of-liquid-in sink) flow-rate)

Figure 3: A model fragment representing fluid flow process between two


containers.

The idea of model fragments presented here is based on the definition presented in [4], which is a
generalization of the representations originally developed by Forbus in his qualitative process theory[2].
An important issue that must be addressed to realize the potential benefits of compositional modeling is
knowledge acquisition. To construct a system that can reason about a variety of problems in a given
domain, one must build up a large library of model fragments covering a substantial portion of the domain.
However, since construction of such a large library is a labor intensive and time-consuming task, efforts are
under way to develop facilities for enabling collaborative construction and reuse of knowledge [5].

2.4. Model formulation and abstraction


For a system intended to aid humans in analyzing a wide variety of the behavior of physical systems in a
given domain, compositional modeling is the most promising approach for automatically formulating a
model. However, in order for the compositional modeling approach to succeed, one must have a
mechanism for selecting model fragments in such a way that the resulting model will be appropriate for the
given analysis problem.

The quality and usefulness of the results of analysis critically depend on the appropriateness of the model.
There are as many possible models of a given subject of study as there are reasons for constructing models.
There is no one "true" or "correct" model since any model is necessarily an abstraction and the
appropriateness of a model depends on one's goal, i.e. what question one wishes to find an answer to by
constructing and analyzing the model.

Appropriateness of a model in this case depends on such factors as whether the model includes relevant
information and whether the model is at the right level of abstraction for answering a given analysis
question. For example, suppose that one is designing an overhead projector. If one wants to know whether
a given design will succeed in projecting an image on the screen, the model should include the optical
aspect of the device being designed. On the other hand, if one is trying to decide whether a fan should be

1 I+ and I- represent influences of one quantity on the rate of change of another. (I+ x y) means that the
phenomenon that has this consequence contributes y to the rate of increase in x. They are examples of
composable equations in . In order to determine the value or the direction of change of an influenced
variable, one must collect all known such influences on the variable and compose them together to
produce a qualitative equation.

7
added to the design to keep the projector from getting too hot, the model that includes the thermal and the
electrical aspects of the device should be used. The question of how to formulate automatically a model
that is appropriate for a given analysis purpose from a body of general knowledge of the domain has
received much attention in the qualitative physics community [6-9].

Closely related to the issue of model formulation is that of model abstraction. If one is given the most
detailed model of the system available to begin with, one way to generate a model that is more appropriate
for a given task is to abstract away some of the irrelevant details in the complex model [10, 11]. Another
way to produce an abstract model is to do so from the simulation results using a somewhat more detailed
model. If the simulated behavior exhibit some pattern (such as repetitive behavior), one could produce a
more abstract model by aggregating the set of model fragments involved in one iteration of the repeated
behavior into one model fragment. An abstract model thus produced can be used to do further prediction at
a coarser grain size than the original simulation [12].

Another approach to model formulation is to infer underlying model from data when no model of the
system is available to begin with. Such techniques can be useful for tasks such as data interpretation,
automatic discovery and diagnosis [13, 14] .

2.5 Other topics


This brief tutorial has so far focused on qualitative reasoning based on qualitative calculus. There are also
other paradigms of qualitative reasoning being pursued in the field. Researchers including Sacks, Yip,
Nishida, and Zhao have investigated techniques based on a method used in the systems dynamic theory.
They have developed techniques for qualitatively characterizing the behavior of complex dynamic systems
using phase portraits of such systems [15-18].

Another type of qualitative reasoning that is an essential part of commonsense reasoning about the physical
world is qualitative spatial reasoning. We certainly seem to reason much about space, shape, and
interactions among solid bodies in everyday life without having precise quantitative data. There has been a
considerable amount of work on automating qualitative reasoning about space and shape [19, 20]. So far, it
has turned out that qualitative spatial reasoning is quite a bit harder than qualitative reasoning about system
dynamics. Forbus et al. have expressed serious doubts about the feasibility of purely qualitative reasoning
about spatial problems, and advocated a combined approach, combining qualitative and quantitative
modeling techniques [19].

A related, difficult problem area is qualitative reasoning about the behavior of distributed-parameter
systems, including problems of weather prediction and modeling diffusion, which are usually modeled with
partial differential equations. Generally, sophisticated finite element analysis programs are used to model
such systems quantitatively. There has been some work on qualitative reasoning about distributed-
parameter systems, but this area is still wide open, mostly unexplored territory [21].

While this tutorial, as well as the articles included in this special track, has focused on qualitative reasoning
about physical domains, the principles developed in the field are equally applicable to non-physical
domains. In fact, the usefulness of qualitative reasoning is even more obvious in other non-physical
disciplines, especially social sciences, where precise quantitative models of the subject are rarely available.
Samuelson discussed qualitative modeling before the field of artificial intelligence existed as such [22].
Researchers in ecology, another field where precise quantitative models are unavailable, also developed
techniques for qualitative modeling of complex systems [23]. Interests are growing in applications of
qualitative reasoning to fields such as finance and economics in recent years [24].

3. The Special Track on Applications of Qualitative Reasoning


This special track attempts to showcase some recent applications of qualitative reasoning techniques to
practical problems. Though these articles represent a small sample of the many exciting efforts under way,
I hope that they help to illustrate the breadth of applications of qualitative reasoning technology to real-
world problems. Five articles are part of this special track: The article by Milne et al. describes TIGER, a
gas turbine monitoring system in use at the Fife Ethylene Plant owned by Exxon Chemical and Shell

8
Chemical. In 1996, TIGER won the European Information Technology Prize, which provides recognition
to innovative information technology products with excellent market potential. A qualitative model of the
turbine system dynamics is used for prediction and diagnosis in TIGER. Producing fielded software
requires making use of not only one but many different areas of technology. In addition to demonstrating
how qualitative reasoning technology can be useful, the article shows how qualitative reasoning fits in with
other components of an overall solution to an important practical problem.

The article by Vescovi et al. describes another monitoring system, this one for monitoring the sintering
process in a steel plant in Brazil. Their system uses a interval-based semi-quantitative model and a
simulation method to monitor the process to achieve safety as well as improved productivity. Interval-
based semi-quantitative simulation lies mid-way between a conventional numerical simulation and purely
qualitative simulation. It allows simulation of models whose functional forms are known but not precise
values of the parameters or the initial values.

The article by Faltings describes a different use of qualitative models for prediction and monitoring. In
order to predict the behavior of dynamic systems for which precise quantitative models do no exist, Faltings
relies on a case library of past cases to perform quantitative prediction but uses qualitative models as
indices into the case library. The paper describes his approach and its application to two different
processes: coffee roasting and decaffination in a Nestle plant.

In the realm of design, the article by Sakao et al. describes an application in generating the software
embedded in a copying machine. In today's electro-mechanical devices, the embedded software for
controlling the device accounts for a significant portion of the cost of design and development of the overall
system. The article describes automatic generation of the control software using a qualitative model of the
hardware component of the system.

One of the most promising application areas of qualitative reasoning is education. It is hoped that
qualitative reasoning systems provide a better basis for constructing a tutorial system for teaching about
physical problems because the level of knowledge used in reasoning and the way solutions are derived --
many qualitative reasoning systems rely on or infer causal dependency relations -- are thought to resemble
the way humans think than numerical simulation systems. The article by Forbus, one of the pioneers of
qualitative reasoning, discusses the advantages of qualitative reasoning for the purpose of education and
describes examples of educational software for teaching thermodynamics being tested in actual classrooms.

In addition to the articles included in the special track, this month's Research Center department (*** note
to the editor *** please put the correct name for the department) features the Model-Based Autonomous
Systems Project team at NASA Ames, developing an intelligent control system for unmanned spacecrafts
using qualitative modeling technology. The Interview department (*** ditto --- I need the correct name
***) features Benjamin Kuipers, who is another one of the pioneers of the qualitative reasoning field. He
developed QSIM, which is one of the first qualitative simulation programs and is widely distributed among
qualitative reasoning researchers. Kuipers leads an active group of researchers on qualitative reasoning at
University of Texas at Austin. Last but not least, this month's Expert Opinion is professed by Ernest Davis
of Courant Institute, who has made a significant contribution to the study of the representation of
commonsense knowledge.

Sources of additional information on qualitative reasoning:


Readings:
Weld, D.S. and Kleer, J.d., ed. Readings in Qualitative Reasoning about Physical
Systems. The Morgan Kaufmann Series in Representation and Reasoning, ed.
R.J. Brachman. 1990, Morgan Kaufmann: San Mateo, CA.
Faltings, B. and Struss, P., ed. Recent Advances in Qualitative Physics. 1992, The
MIT Press.
Kuipers, B., Qualitative Reasoning: Modeling and simulation with incomplete
knowledge. 1994, MIT Press.

9
Forbus, K., Qualitative Reasoning, in Handbook of Computer Science. 1997, CRC
Press. Forthcoming.

Workshops
International Workshop on Qualitative Reasoning
An international workshop which has been held every years since 1987. Contact Dr. Liliana Ironi
(liliana@supers1.ian.pv.cnr.it) for more information about this year's workshop being held in Cortona,
Italy. Contact Prof. Feng Zhao (fz@cis.ohio-state.edu) for more information about the workshop in 1998.

Practical Applications of Qualitative Reasoning


A day-long workshop with a focus on applications. It has been held in conjunction with the last three
IJCAI including the one this year. Contact Prof. Gautam Biswas (biswas@vuse.vanderbilt.edu) for
information about this year's workshop.

Web-site
http://ai-www.aist-nara.ac.jp/doc/qphysics/.
Home page of the international QR community with links to other related sites

References
1. Kuipers, B., Qualitative Simulation. Artificial Intelligence, 1986. 29.

2. Forbus, K.D., Qualitative Process Theory. Artificial Intelligence, 1984. 24(1-


3).

3. Forbus, K.D., The Qualitative Process Engine, in Readings in Qualitative


Reasoning about Physical Systems, D.S. Weld de Kleer, J., Editor. 1989, Morgan
Kaufmann:

4. Bobrow, D., et al. A Compositional Modeling Language. In Proceedings of The


Tenth International Workshop on Qualitative Reasoning. 1996. Fallenleaf Lake,
CA. AAAI Press.

5. Iwasaki, Y., et al. A web-based compositional modeling system for sharing of


physical knowledge. Submitted to Proceedings of The Fifteenth International
Joint Conference on Artificial Intelligence. 1997. Nagoya, Japan.

6. Falkenhainer, B. and Forbus, K., Compositional modeling: finding the right


model for the job. Artificial Intelligence, 1991. 51(1-3).

7. Iwasaki, Y. and Levy, A.Y. Automated Model Selection for Simulation. In


Proceedings of The Twelfth National Conference on Artificial Intelligence.
1994. Seattle, WA. The AAAI Press.

8. Nayak, P., Joskowicz, L., and Addanki, S. Automated Model Selection Using
Context Dependent Behaviors. In Proceedings of Tenth National Conference on
Artificial Intelligence. 1992. San Jose, CA. The AAAI Press/The MIT Press.

9. Rickel, J. and Porter, B. Automated Modeling for Answering Prediction


Questions: Selecting the Time Scale and System Boundary. In Proceedings of
The Twelfth National Conference on Artificial Intelligence. 1994. Seattle, WA.
The AAAI Press.

10
10. Williams, B.C. Capturing How Things Work: Constructing Critical Abstraction
of Local Interactions. In Proceedings of The AAAI Workshop on Automatic
Generation of Approximations and Abstractions. 1990.

11. Weld, D.S. Approximation Reformulation. In Proceedings of The Eighth


National Conference on Artificial Intelligence. 1990. Boston, MA. AAAI Press/
The MIT Press.

12. Weld, D.S., The User of Aggregation in Causal Simulation. Artificial


Intelligence, 1986. 30(1): p. 1-17.

13. Forbus, K.D. Interpreting measurements of physical systems. In Proceedings of


The Fifth National Conference on Artificial Intelligence. 1986. Philadelphia,
PA. Morgan Kaufmann Publishers.

14. Ramachandran, S., Mooney, R.J., and Kuipers, B.J. Learning Qualitative Models
for Systems with Multiple Operating Regions. In Proceedings of The Eighth
International Workshop on Qualitative Reasoning about Physical Systems.
1994. Nara, Japan.

15. Sacks, E., Automatic qualitative analysis of dynamic systems using piecewise
linear approximations. Artificial Intelligence, 1990. 41: p. 313-364.

16. Yip, K., KAM: A system for intelligently guiding numerical experimentation
by computer. 1991, The MIT Press.

17. Nishida, T., Mizutani, K., Kubota, A. and Doshita, S. Automated Phase Portrait
Analysis by Integrating Qualitative and Quantitative Analysis. In Proceedings
of The 10th National Conference on Artificial Intelligence. 1991. Boston, MA.

18. Zhao, F., Extracting and representing behaviors of complex systems in phase
space. Artificial Intelligence, 1994. 69: p. 51-92.

19. Forbus, K., Nielsen, P., and Faltings, B., Qualitative Spatial Reasoning: The
CLOCK project. Artificial Intelligence, 1991. 51(1-3): p. 417-471.

20. Joskowicz, L. and Sacks, E.P., Computational kinematics. Artificial


Intelligence, 1991. 51(1-3): p. 381-416.

21. Lundell, M. A qualitative model of physical fields. In Proceedings of The tenth


International workshop on qualitative reasoning. 1996. Fallenleaf Lake, CA.
AAAI Press.

22. Samulelson, P.A., Foundations of Economic Analysis. 1947, Cambridge,


Massachusetts: Harvard University Press.

23. Puccia, C.J. and Levins, R., Qualitative Modeling of Complex Systems. 1985,
Harvard University Press.

24. Hamscher, W., Kiang, M., and Lang, R., Qualitative Reasoning in Business,
Finance and Economics: Introduction to the Special Issue. Decision Support
Systems, 1995. 15.

11

You might also like