Proceedings of the 1991 Winter Simulation Conference

Barry L. Nelson, W. David Ke]ton, Gordon M. Clark (eds.)

QUALITATIVE MODELING AND SIMULATION: PROMISE OR ILLUSION

Fran~ois E, Cellier

Department of Electrical and Computer Engineering

University of Arizona
Tucson, Arizona 85721

ABSTRACT state variables change their values infinitely often.”

No other mathematical model shares this property.
In this panel discussion, questions shall be addressed Continuous-time models are represented through
relating to the potential usefulness of qualitative sets of differential equations. Among the continuous-
modeling and simulation. When might qualitative time models, two separate classes can be distin-
versus quantitative modeling/simulation be justified? guished: the lumped parameter models, which are de-
What types of qualitative models if any are useful scribed by ordinary di#erential equations (ODES), in
and under which conditions? What can qualitative general:
simulations reveal that quantitative simulations can-
not ? Under what conditions are combined quantit a-
tive/qualitative models feasible/meaningful? x = f(x, u, t) (1)

In this article, some basic definitions are presented

and for the special case of linear system%
that may serve as a basis for the discussion. After all,
we must first agree on some common ground before
particular properties can be explored. x= Ax+ Bu (2)

and the distributed parameter models, which are de-

1 TYPES OF MATHEMATICAL MODELS
scribed by partial differential equations (PDEs) such
What types of mathematical models do exist? A first as the diffusion equation:
category is the set of continuous-time models. Figure
1 shows how a state variable a changes over time in au
—=
(Pu
(3)
a continuous-time model. a r“~

The second class of mathematical models to be

Continuous-Time Model mentioned is the set of discrete-time models. Figure 2
depicts the trajectory behavior exhibited by discrete-
‘~
time models.
10
x
5

o
0.0 0.5 1.0 1.5 20 25 3.0 3.5 4.0 4.5 5.0
Time

Figure 1: Trajectory behavior of continuous models

0.0 0.5 1.0 1.5 20 25 3.0 3.5 4.0 4.5 5.0

Time
We can give the following definition for continuous-
time models: “Continuous-time models are character-
ized by the fact that, within a finite time span, the Figure 2: ‘Trajectory behavior of discrete-time models

1086
 Qualitative Modeling and Simulation 1087

In this type of models, the time axis is dis- ticular time instants are seen as noteworthy, namely
cretized. Discrete-time models are commonly repre- those where a state change occurs.
sented through sets of difference equationa, at least if Notice that nothing is said about the nature
the discretization is equidistantly spaced. Such mod- of the state variables themselves. The three
els can be represented as: model/simulation types differ only in their interpreta-
tion of time. If the state variables themselves are real
Xk+l = f(x~,u~,tk) (4) valued, the model is called quantitative, otherwise it
is called qualitative. Figure 4 shows the trajectory
The third and final class of models is the set of behavior of a qualitative model.
discrete-event models. Paradoxically, the time axis
of discrete-event models is usually “continuous” (i.e.,
reai rather than integer), but discrete-event models
differ from the continuous-time models by the fact
that, in a finite time span, only a finite number of
state changes may occur. Figure 3 depicts the typical
trajectory behavior of a state variable in a discrete-
event simulation.

Discret&Event Model
ly,
I

Figure 4: Trajectory behavior of qualitative models

Different types of qualitative models differ in the

manner in which the state variables are discretized,
0.0 0.5 1.0 1.5 20 2.5 3.0 3.5 4.0 4.5 5.0
and in the manner in which time is advanced.
Time

2 SOME USEFUL DEFINITIONS

Figure 3: Trajectory behavior of discrete-event models

The following terminology shall be used:

Discrete-event models are usually described by an 1. Qualitative variables are variables that assume
enumeration of all possible event types together with a finite ordered set of qualitative va’lues, such
either a list of times when these events occur (a so- aa “minuscule,” “small: “average,” “limge,” and
called event calendar) or a set of conditions under “gigant it.” The lit erature on quantitative soft
which they occur (activity scanning). sciences is a little more precise on this definition
The three types of models differ in their interpre- than the literature on artificial intelligence. For
tation of time. In continuous-time simulations, time instance, Babbie (1989) distinguishes between:
is (at least conceptually) an analog variable. Digi-
(a) Nominal measures, i.e., variables whose val-
tal continuous-time simulations proceed by advanc-
ues have the only characteristics of exhaus-
ing the simulation clock in sufficiently small steps
tiveness and mutual exclusiveness. Nomin-
so that the human observer of the simulation results
al measures are unordered sets. Typical
is seduced into smoothing out the trajectory behav-
nominal variables might be the religious af-
ior, i.e., he or she does not perceive the (on a digi-
filiation, or the hair color of a person. Such
tal computer necessary) time quantization aa an es-
variables are not useful as state-variables in
sential or noteworthy property of his or her model.
a simulation. They can play a role as pa-
Discrete-time simulations perceive time as an integer
rameters.
or fixed point variable. Time is advanced by a fixed
clock increment that is sufficiently large to make it (b) Ordinal measures, i.e., variables that are
essential/noteworthy. Discrete-event simulations fi- nominal, and in addition, are rank-ordered.
nally proceed from one event time to the next, i.e., These variables are what I called above
while time is perceived aa a real variable, only par- qualitative variables. However, sometimes
1088 Cellier

we shall let go of the condition of mutual around that minimize ambiguity in a systematic fash-
exclusiveness, for example, when we oper- ion? I don’t know of any such techniques. Hopefully,
ate on fuzzy sets. one of the panel members will address this important
issue.
(c) Interval measures, i.e., variables that are or-
dinal, and in addition, have the property
that a distance measure can be defined be- 3 TYPES OF QUALITATIVE MODELS
t ween any two values, that is: interval vari-
ables can be added to and/or subtracted 3.1 Naive Physics Models
from each other. A typical candidate for
a “soft” interval variable might be the in- Naive physics encompasses a set of different tech-
telligence quotient. niques for knowledge-based reasoning about physi-
cal systems. Different representatives of this type of
(d) Ratio measures, i.e., variables that are in-
qualitative models are described in Bobrow (1985).
terval measures, and in addition, have a
All naive physics models have in common that they
true zero point.
discretize the state space in a very crude way, namely
2. Qualitative behavior denotes a time-ordered set the set { – O + }. A state variable is characterized
of values of a qurdit ative variable, i.e., an episode. as either negative, —, zero, O, or positive, +. Some
Episodes are qualitative trajectories. researchers view these as three different states, while
others view – and + as two different states with O
3. Qualitative models are models that operate on being a “landmark” separating the two states. Both
qualitative states. interpretations are meaningful. The former uses O to
represent any state that is sufficiently small in mag-
4. A qualitative simulation is an episode generator
nitude, whereas the latter interprets O as a transitory
that infers qualitative behavior from a qualita-
condition that separates the states – and +.
tive model.
The legitimation for this methodology is taken from
Qualitative state variables are frequently ordinal the observation that, in a mechanical system, the di-
measures, i.e., no dist ante information is preserved rection of forces seems to carry more information than
between neighboring states. In this respect, Fig.4 is the magnitude. We don’t need to know the magni-
atypical. Also, time in a qualitative simulation is of- tude of the gravitational force to conclude that all
ten perceived as a qualitative variable as well. One free-moving objects will eventually fall towards the
“unit” of qualitative time is the time that elapses be- gravitational center.
tween two consecutive state changes. Obviously, since Some researchers have observed that, while the
state changes don’t have to be equidistantly spaced, magnitude of the state variables may still be impor-
a “unit” of qualitative time is not a quantum. While tant to qualitatively describe the behavior of a sys-
qualitative models with quantitative time can be rep- tem, the magnitude of their time derivatives may no
resented through integer-state discrete-event models, longer be important. Consequently, some researchers
qualitative models wit h qualitative time could be rep- describe their models in terms of derivatives only
resented through integer-state discrete-time models. (qualitative derivatives are sometimes referred to as
However, a more commonly used representation for confluences). Others extended the methodology by
such models is the finite state machine. allowing additional levels and landmarks for the state
While quantitative simulation always denotes the variables while restricting state derivatives to the
trajectory behavior of a quantitative model in re- original ternary set.
sponse to a particular expem”ment, qualitative simu- Some researchers describe their naive physics
lation often aspires to describe the episodical behav- models through qualitative stat e equations, others
ior of a qualitative model in response to all possi- through a rule base (a constraint set relating qual-
ble experiments. Consequently, while the result of a itative variables to each other). Morgan (1990) de-
quantitative simulation is one particular trajectory scribes an automated procedure to translate a qual-
behavior, the result of a qualitative simulation may itative state-space model into a finite state machine
be a (possibly extensive) set of all feasible episodi- (i.e., a special type of a rule base). Many of these re-
cal behaviors. It is desirable to describe a qualitative searchers concentrate more on the modeling than on
model in such a way as to minimize the cardinality of the simulation aspects, i.e., for them the generation
this set. However, this is often not possible, and it is of the qualitative model is more important than the
therefore quite common that a qualitative simulation generation of episodical behavior. Those that gener-
drowns in the sea of ambiguity. Are there techniques ate episodical behavior usually operate on qualitative
Qualitative Modeling and Simulation 1089

rather than quantitative time. of event times, it could be that one event precedes
another or vice-versa. Symbolic discrete-event mod-

3.2 Inductive Reasoning Models els allow us to fomulate event times as polynomials
of time with unknown or partially known (possibly
Inductive reasoning describes a variety of pattern- fuzzy) coefficients. Symbolic discrete-event simula-
bssed techniques to reason about relations between tion generates all trajectories that are feasible due
qualitative variables. While naive physics models
to the fuzziness of these parameters. This technique
strive to capture the structure of models, inductive
has been described in Zeigler and Chi (19!? 1). The
reasoning models aspire to capture the behavior di- fuzziness of event times is expressed through so-called
rectly. Consequently, inductive reasoners generate time windows. The time window in formaticm is sub-
directly and immediately finite state machine repre-
sequently also used for the purpose of fault diagnosis.
sentations of systems, whereas naive physics models
Symbolic discrete-event models provide us with a
arrive at such a representation only indirectly if at
second knowledge-based approach to reason about
all. A good overview of various inductive reasoning
qualitatively known systems.
approaches is given in Klir (1985).
Inductive reasoners are based on time quantization
in addition to state quantization. Consequently, they 3.4 Neural Network Models
operate on quantitative rather than qualitative time.
Neural networks provide us with a second pattern-
Inductive reasoning models are optimization models.
based approach to qualitative (and even quantita-
They optimize the forecasting power of the model by
tive) modeling. Contrary to the inductive reasoners
minimizing the indet erminism or ambiguity of state
discussed earlier, the number of discrete states of a
transitions. Cellier (1991) describes techniques for neural network can be quite large. Some neural net-
optimizing the state and time discretizations. How-
works (such as backpropagation networks) even oper-
ever, it is equally important to select a representative
ate on a continuous state-space, i.e., can be used for
set of state variables to base the model upon. I don’t the identification of quantitative models. However,
know of any technique that would allow us to system- such models cannot be used to enumerate all possible
atically determine an optimal set of state variables. system behaviors, and no information is generated as
Hopefully, one of the panel members will address this to the likelihood of a particular behavioral pattern.
important issue.
Yet, neural network models can be very powerful tools
Contrary to the finite state representations gener- for capturing the behavior of (partially) unknown sys-
ated from nafve physics models, inductive reasoners tems. An overview of this technique is given in Cellier
contain information about the likelihood of any par- (1991).
ticular state transition. This is important for model
validation purposes. If the accumulated likelihood of
4 COMBINED QUANTITATIVE AND
a particular episode drops below a level that can be
QUALITATIVE MODELS
user-specified, forecasting will come to a halt. There-
fore, the user can guarantee that his or her model
It is quite common that properties of a system are
will not forecast behavior beyond a time for which
partially known and partially unknown. It is there-
the available data are insufficient to substantiate the
fore desirable if available knowledge about a system
prediction.
can be coded into the model while unknc~wn prop-
Similar to the naive physics models, it is possible
erties can be treated in a qualitative manner. This
to use inductive reasoners to enumerate all possible
justifies the call for combined quantitative and qual-
(or likely) system behaviors.
itative models. However, the generation of combined
models is basically virgin territory. Very little has
3.3 Symbolic Discrete-Event Models
been writ ten about such combined models. Obvi-
As mentioned earlier, discrete-event models are ously, the discrete-event approach lends itself natu-
equally well suited to represent qualitative as quanti- rally to such models. However, this approach requires
tative models. In the qualitative case, the event cal- that the model as a whole be formulated as a discrete-
endar must contain the time instants when the system event model. This is not always desirable. Inductive
transits from one qualitative state to another. How- reasoners could be integrated with continuous-time
ever, as qualitative states may be fuzzy, also state models using the fuzzy measure approach introduced
transitions will be unsharp, and therefore also event by Li and Cellier (1990). However, no practical expe-
times. Ambiguity can occur since a unique sequence rience is available yet. It is not clear to me how naive
of events may no longer be given. Due to the fuzziness physics models could be combined with quantitative
1090 Cellier

models. I hope that one or the other of the panel Dr. Cellier’s main scientific interests concern mod-
members will address the issue of how which type of eling and simulation methodology, and the design of
qualitative models can be combined with quantitative advanced software systems for simulation, computer-
models. aided modeling, and computer-aided design. He has
designed and implemented the GASP–V simulation
package, and he was the designer of the COSY simu-
5 CRITICAL ASSESSMENT
lation language a modified version of which under the
name of SYSMOD has meanwhile become a standard
While there exists a good amount of literature already
by the British Ministry of Defence. Dr. Collier has au-
describing various methodologies for qualitative mod-
thored or co-authored more than fifty technical pub-
eling (as described above), very few references discuss
lications, and he just published his first textbook on
practical real-life applications of these techniques (ap-
continuous system modeling. He served as a chairman
plications that reach beyond simple schoolbook ex-
of the National Organizing Committee (NOC) of the
amples). It is not clear how well this technique scales
Simulation’75 conference, and as a chairman of the
up to solve realistically large problems. I hope that
International Program Committee (IPC) of the Sim-
one or the other of the panel members will discuss
ulation’77 and Simulation’80 conferences, and he has
practical experiences with qualitative modeling and
also participated in several other NOC’S and IPC’S.
simulation.
He is associate editor of several simulation related
journals, and he served as vice-chairman oft wo com-
REFERENCES mit tees on standardization of simulation and model-
ing software.
Babbie, E. 1989. The Practice of Social Research,
5th edition, Wadsworth Publishing Company, Bel-
mont, Calif.
Bobrow, D.G. 1985. Qualitative Reasoning About
Physical Systems, MIT Press, Cambridge, Mass.
Cellier, F.E. 1991. Continuous System Modeling,
Springer-Verlag, New York.
Klir, G.J. 1985. Architecture of Systems Problem
Solving, Plenum Press, New York.
Li, D. and F.E. Cellier. 1990. “Fuzzy Measures in In-
ductive Reasoning,” Proceedings 1990 Winter Sim-
ulation Conference, New Orleans, La., pp. 527–
538.
Morgan, A.J. 1990. “Accuracy in Qualitative De-
scriptions of Behavior,” Proceedings 1990 WinteT
Simulation Conference, New Orleans, La., pp. 520-
526.
Zeigler, B.P. and S.D. Chi. 1991. “Symbolic Discrete
Event System Specification,” Proceedings IEEE
Conference on AI, Simulation and Planning in
High Autonomy Systems, Cocoa Beach, Fla., pp.
130-141.

AUTHOR BIOGRAPHY

FRANCOIS E. CELLIER received his B.S. degree

in Electrical Engineering from the Swiss Federal Insti-
tute of Technology (ETH) Zurich in 1972, his M.S. de-
gree in Automatic Control in 1973, and his Ph.D. de-
gree in Technical Sciences in 1979, all from the same
university. Following his Ph. D., Dr. Cellier worked
as a Lecturer at ETH Zurich. He joined the Uni-
versity of Arizona in 1984 as an Associate Professor.