You are on page 1of 8

MODELING VISUAL REASONING ON GRAPH USING

QUALITATIVE INTERPRETATION
S.M.F.D. Syed Mustapha
Faculty of Computer Science and Information Technology
University of Malaya, Lembah Pantai
50603, Kuala Lumpur
Tel:+(603)7560022 ext 6346 Fax: +(603)7579249 e-mail:symalek@fsktm.um.edu.my

Abstract: Human visual reasoning and understanding are important in the study of human cognition
(Pisan, 1995). It is a constituent discipline in the field of artificial intelligence. Such research study
advocates to various orientations such as image processing, neural networks, parallel distributed
processing, machine learning and others. Human graph recognition through subjective visualization
substantially plays the role in various applications such as engineering, fluid mechanics and education.
Particularly, in reasoning towards the behavior of physical relations and processes via the piece-wise
graph representation. In this paper we discuss an artificial intelligence technique so-called qualitative
reasoning which is used to generate qualitative interpretation in analyzing graphs. Selection on graph
models can be misleading and require qualitative understanding about the nature of the graph. We
define briefly the scope of qualitative reasoning, discuss the reasoning techniques by providing several
selected research works as case studies.

1.. INTRODUCTION
Graph expresses relationships between variables in a pictorial form. It serves both as devices to
aid in the representation of obscured raw data and make holding information in the human’s
memory easier (Pisan, 1994). Different types of graphs indicate distinctive refinements during the
process of recognition. For instance, in pie chart, the relative size of portions are observed, bar
graphs show the relative height or amount, scatter plots display the pattern of the data and line
graphs portray the continuous relationship between two variables.

The graph recognition on the line graph is our main interest because of its usage in various
applications are far-reaching. Human being recognizes a graph through subjective visual
inspection. Interpretation of the graph is done by giving qualitative description based on the
physical appearance of the slope, direction, continuity and spatial relationship. A monotonic x-y
relationship commonly described as linear model or non-linear model if it shows a continuous
change in slope and direction. Nomenclatures such as the parabolic, the hyperbolic and the
sinusoidal wave posses qualitative term which are frequently referred by human. For anonymous
graph, separate parts of the graph have to be labeled individually so that a single-named reference
can be made by concatenating the labeled parts. The application of such technique has been
shown in (McIllraith, 1989; Ritter and Will, 1990).

In this paper, we describe an approach of artificial intelligence, so-called qualitative reasoning. It


is a technique incorporated into a computer system such that interpretation of a piece-wise linear
graph can be automated. The idea of the qualitative reasoning are introduced in section 2,
followed by its implementation in several areas such as in well-test data interpretation in section
3 and modeling the fluid behavior in section 4.

2.QUALITATIVE REASONING
The notion of qualitative reasoning emerges in the realm of physics (Forbus, 1988). The
objectives are to represent and reason about the physical behavior of a system. Professor Kenneth
Forbus (Forbus, 1998) and his research group are the founder of this theory. They expounded this
concept with an example of an object attached to a spring sliding through a friction-free surface
as illustrated in Fig. 1 (Cohn 1989).
2

Spring Object

Figure. 1 A sliding object on a friction-free


surface
If we posed a question “what would happen if the extended spring of the object is released from
position 2?”. Without a need to engendering any mathematical expression such as F = kx, human
being qualitatively explain “the object oscillates back and forth several times for a short time
until it eventually stops”. A simple conjecture of this has been employed extensively in various
simple to complex areas such as modeling the coffee machine problem (Bonissone and Valvanis,
1985), automotive circuit simulation (Snooke and Price, 1997) and reasoning towards the
structural behavior of fluids (Yip, 1995). Several organizational structures, so-called ontology,
have been introduced. Among the prominent ones are contraint based (Kuipers, 1986),
component centered (Kleer and Brown, 1984) and process centered (Forbus, 1988).

In this paper, we focus the implementation of the qualitative reasoning in graph recognition and
understanding. We also mention the essential of this technique in steering to the selection of
correct model. The dominant properties of qualitative reasoning over the other reasoning methods
such as rule-based reasoning and expert system are:
I. The ability of performing reasoning on the behavioral of physical system without the
necessity of building up a precise model (Cohn, 1989).
II. Reduce the computational time cost on many occasions where approximate results are
superior to exactness.
III. Exploring the symbolic modeling paradigm, so-called qualitative model, which
reasons from the fundamental theories and incomplete data or knowledge. This allows
the generation of qualitative simulation (Kuipers, 1986).
IV.Simplify the parameter setting and model selection of quantitative model. This
reduces the searching space and selection complexities (Mustapha, 1997, 1998).
The graph recognition process undertakes two main stages namely, qualitative modeling and
quantitative modeling. The qualitative modeling, which resorts to qualitative reasoning technique,
captures the behavioral of the graph and its physical features. At this stage also, the class of
model library can be determined. This can reduce computational cost by narrowing the range of
possible graphs to more accurate ones (Capelo et al., 1991, 1992, 1993, 1995, 1996). In the
quantitative modeling stage, parameter values are estimated. The accuracy of this work depends
on the well-chosen graph model from the early stage. In the next section we review the
application of this technique in interpreting graph displaying an actual well pressure build up
data.

3.WELL-TEST DATA INTERPRETATION


Two relevant cases on the well-test data interpretation are discussed here. In the first example,
Ritter and Will (Ritter and Will, 1990) showed an interpretation system assembled with the aid of
an expert system shell complimented with graphical displays of an actual well-pressure build up
data. The transient pressure analysis is essential for a good estimation of permeability formation,
well damage and average pressure in the drainage volume of the well.

Figure 2 shows a diagnostic model produced by the interpretation system. The small triangular
shapes are the real well-test data. The continuous lines are the master curves, which are split into
three period of time, early time, middle time and late time.

User input an appropriate names (some examples are given in Table 1) to the expert system. The
authors in (Ritter and Will, 1990) claimed that there is no need to automate the labeling process,
as the task is obvious and trivial to users and experts. The rule-based expert system will perform
reasoning to select suitable model solution such as Homogeneous, Dual_Porosity,
Frac_Layer_Fault and Frac_Homo_Closed.

EARLY TIME MIDDLE TIME LATE TIME

UnitSlope

HalfSlope

QuarterSlope

Figure 2. Diagnostic Model (after Ritter and Will)

EarlyTime MiddleTime LateTime Model Solution

Missing Stabilized MissingHomogeneous


UnitSlope FlatMinimum Stabilized Dual_Porosity
HalfSlope SmoothMinimum Increasing Frac_Layer_Fault
QuarterSlope SharpMinimum FallingOff Frac_Homo_Closed

Table 2. Model solution for time periods

We elaborate the work of McIllraith (McIllraith, 1989) as second example. Unlike Ritter and
Will, labeling the components of a curve in her work is fully automated. The adoption of
qualitative reasoning in the graph recognition is shown to be an effective method. The selection
of possible quantitative model is augmented by the knowledge of physical appearance of the
graph, prioritizing the region to be assessed and classifying the data sets to a smaller range of
possible graph models. The diverse sources of knowledge can be incorporated into an
interpretation system. The graph is described using qualitative description language, which is
represented using BNF notation. Some of these are given below:
<graph> ::= <region><graph>
<graph> ::= <region>
<region> ::= <length><region-type>
<region-type> ::= <curve descriptor><curve>
<region-type> ::= <slope grade><line>
<curve descriptor> ::= <curve rate><curve shape>
<curve rate> ::= <gradual> | <rapid>
<curve shape> ::= <concave> | <convex>
<slope grade> ::= <gentle> | <moderate> | <steep>
<length> ::= <short> | <average> | <long>
A graph is composed of a region and another graph. The region is the important part of a graph in
which its features are essential to be assessed. A region can be an entire graph or part of a graph.
The length and region-type characterize the region. The attributes such as gradualness, rapidness,
concavity characterized the curve descriptor, which are also component types of a region-type.

There are four levels of visual interpretation introduced by McIllraith (McIllraith, 1989). They
are namely, regionalization, labeling, smoothing and curve characterization. In the regionalization
stage, curve discontinuity and abrupt change of direction is detected. The curve is then
partitioned into several segments. In labeling, each segmented curve is mathematically
characterized by fitting to low-order polynomial function. Each region is mathematically
described in terms of shape and length (McIllraith, 1989). Symbolic pattern language
characterization mentioned earlier is used to map the mathematical characterization to qualitative
description language. The smoothing performs subjective smoothing to detect anomalies behavior
of a graph. The occurrences of spikes, outliers or a line between two curves are repaired. There is
a need to re-label if mislabeled has been found. The final stage of visual interpretation is curve
characterization, which is executed in two different manners. The first is the identification of
characteristic curve forms in the observed data. This leads to the production of physical features
such as positive skin, wellbore storage and partial penetration. This can be done by performing
curve description parsing. For example, the following qualitative curve description,

AVERAGE GRADUAL CONCAVE CURVE + LONG GRADUAL CONVEX SHAPE

can produce the following physical features

POSITIVE SKIN
WELLBORE STORAGE
PARTIAL PENETRATION.

The three interpreted physical features above indicate that at least three rules have matched the
curve description. The second type of curve characterization is determination of statistical model.
The two of these models are Gringarten SAS model and Agarwal SAS model. Each model
displays a set of master curves. The observed data is matched to one of these curves. The first
stage of curve matching is the symbolic matching of the qualitative shape of the observed data to
the ones in master curves. Heuristic knowledge is applied to decide which part of the region is to
give priority. The set of master curves of the statistical model is reduced to several potential
models. The second stage involves more stringent mathematical matching. The weighted least
squares fit is used to guide the selection of best-matching curve.
The two examples mentioned above proved the necessity of qualitative reasoning. Human’s
subjective visual inspection on the graph can be automated using this technique.

4.MODELING FLUID BEHAVIOR


In the field of fluid mechanics and modern study of non-Newtonian fluid such as in rheology
(Barnes et. al, 1989), numerical modeling is essential in estimating accurate parameters. Before
selecting the appropriate accurate models, experimental data are observed and parts of the graph
are analyzed qualitatively. There are several pronged-reasons for doing this, among them are:
1. To analyze graph components to detect physical response such as elasticity, viscous,
recoverable or irrecoverable deformation.
2. To determine the relevant class of constitutive equations.
3. To eliminate inappropriate models through qualitative observation on regions of a
graph.
4. To refrain ill-posed problem (Winter and Baumgaertel, 1993) and evaluation on
misleading models. These models may be proven mathematically valid but can be
unphysical.

Figure 3. An anatomy of creep test graph model (strain vs. time)


We explain the first two reasons with an example. Figure 3 shows a graph model of a creep test.
A creep test is a process of applying stress on a fluid sample and measured the strain in a
cessation of stress (Ferguson, 1995). Segmenting the graph into several components namely OA,
AB, BC, CD, DE and EF enables the analysis of a graph. The strain responses, t0 – t1, are the
instantaneous elastic deformation, delayed elastic strain, viscous and labeled as OA, AB, and BC
respectively. This is followed by cessation of stress, t1 – t, where CD is the instantaneous elastic
deformation, DE is partial recoverable deformation caused by the delayed elastic strain (AB)
during the imposition of strain and EF shows irrecoverable deformation due to linearity at BC.
Qualitative observation can be made by analyzing the general properties such as linearity,
concavity, vertical or horizontal line and asymptotic. The general properties can be used to
determine the physical responses of the graph. The vertical lines OA or CD indicate the existence
of an instantaneous deformation. The concavity of AB or DE exhibits the delayed response. The
linear line of BC or the non-asymptotic EF means an irrecoverable deformation. The qualitative
behavior can be represented in logical triplet as
QB = (α, β, γ),
where α, β and γ respectively are instantaneous elasticity, delayed elasticity and irrecoverable
deformation. These symbolic representations accept, T, for existence and, F, for non-existence of
the physical responses. The logical triplets can lead to any one of the four admissible classes of
constitutive equations,

m m
(T , T , T ) ⇒ ∑
i=0
D s =
i

i= 0
D ie

m m +1
( F ,T ,T ) ⇒ ∑
i=0
D is = ∑
i =1
D ie

m m +1
( F ,T , F ) ⇒ ∑i= 0
D is = ∑i=0
D ie

m +1 m +1
(T , T , T ) ⇒ ∑
i=0
D is = ∑
i =1
D i e.

It is not the intention to flaunt the elegance of the equations above but rather to emphasize that
there are numbers of possible quantitative models, which have been categorized into four classes.
The amount of computational time and problem-state space for an algorithm to traverse in
determination of an accurate model from these possibilities is unattainable. The knowledge of
existence and non-existence of the physical responses will help to reduce the searching to a single
class. This can be achieved by the recognition of physical appearance of the graph.

First Newtonian region

Power Law region

Second Newtonian region

Figure 4. Power Law model and Carraeu-Yasuda


model

The last two reasons explain that purely statistical model-fitting is able to produce numerical
values with the lowest least-squares errors. However, parameter values such as negativity and
infinity do not have any physical meaning. The selection of appropriate models has to give
priority to physical consideration. Figure 4 compares two viscosity models, Power Law and
Carraeu-Yasuda models. Both models can describe mathematically the three regions, namely,
First Newtonian region, Power Law region and Second Newtonian region. However, for the
Power Law, this is not physically possible. Power Law model is given as,

τ = kγ n , since τ = ηγ


kγ n k
η= = 1− n
γ γ
where τ is the shear stress, γ is the shear rate, η is the viscosity ,
k is the consistenc y and n is the Power law index.
When n < 1 (for shear thinning fluid), η approaches ∞ which is impossible for fluids. At a very
high shear rate, the Power Law model also fails when the viscosity eventually reaches constant.
Qualitative reasoning is applied to find the appearance of data at the plateau regions and avoid
them from fitting to the Power Law model. Instead, other models such as Carraeu-Yasuda model
is used.

5. CONCLUSION
We briefly described the application of qualitative reasoning in the areas of graph recognition
and understanding. The impacts of such studies have led to several significant usage such as in
education. Graphs have become essential visual aids in explaining the nature of a business,
engineering process, economics, mathematics and many others to students. The qualitative
interpretation and description of a graph lends itself in generating discussion using natural
language. This has opened avenues to new teaching paradigm, so-called, intelligent tutoring
system. This contradicts to the traditional mathematical descriptions of using figures and
primitive equations. Qualitative interpretation also allows descriptor labeling to be input
automatically to other reasoning system such as the expert system in which was manually done
previously.

REFERENCES

Barnes, H.A.., Hutton, J., and Walters, K. (1989). Introduction to rheology, Elsevier Science
Publishing Company, Inc., Amsterdam.
Bonissone, P.P. and Valvanis, K.P. (1985). “A comparative study of different approaches to
qualitative physics theories”, Second AI Applications Conference – IEEE, 236- 243.
Capelo, A.C., Ironi, L., and Tentoni, S. (1991). “An algorithm for automated generation of
rheological models”, Applications of Artificial Intelligence in Engineering VI, 963-979.
Capelo, A.C., Ironi, L., and Tentoni, S. (1992). “A qualitative simulation algorithm for the
rheological behavior of viscoelastic materials”, Applications of Artificial Intelligence in
Engineering VII, 1117-1130.
Capelo, A.C., Ironi, L., and Tentoni, S. (1993). “A model-based system for the classification and
analysis of materials”, Intelligent Systems Engineering Autumn, 2(3):145-158.
Capelo, A.C., Ironi, L., and Tentoni, S. (1995). “Automatic selection of an accurate model of a
viscoelastic material”, Ninth International Workshop on Qualitative Reasoning about physical
system, 32-43.
Capelo, A.C., Ironi, L., and Tentoni, S. (1996). “The need for qualitative reasoning in automated
modeling:a case study”, The Tenth International Workshop on Qualitative Reasoning, 32-39.
Cohn, A.G. (1989). “Approaches to qualitative reasoning”, Artificial Intelligence Review, (3),
177-232.
Ferguson, F. (1995). “Application of rheology to polymer processing”, In Rheological
Fundamentals of Polymer Processing, Kluwer Academic Publishers.
Forbus, K. (1988). “Qualitative physics: past, present and future”, Exploring Artificial
Intelligence, ed. Howard Shrobe, Morgan Kaufmann Publishers, Inc., 239-296.
Forbus, K. (1998). “Kenneth Forbus’s Home Page”, http: // www.ils.nwu.edu /~e_for_e/people/
forbus.html
Kuipers, B. (1986). “Qualitative simulation”, Artificial Intelligence, 29:280-388.
Kleer, J. de and Brown, J.S. (1984). “A qualitative physics based on confluences”, Artificial
Intelligence, 24:7-83.
Mcilraith, S.A. (1989). “Qualitative data modeling: Application of a mechanism for interpreting
graphical data”, Computational Intelligence, 5(2):111-120.
Pisan, Y. (1995). “A visual routines based model of graph understanding”, In Proceedings of
the Seventeenth Annual Conference of the Cognitive Science Society, Hillsdale, NJ:
Lawrence
Erlbaum Associates, 692-697.
Pisan, Y. (1994). “Visual reasoning about physical properties via graphs”, The Eighth
International Workshop on Qualitative Reasoning about Physical Systems.
Ritter, S.R. and Will, R.P. (1990). “Enhanced well test interpretation with an Expert System”,
1990 Conference on Artificial Intelligence in Petroleum Exploration and Production, 25-32.
Snooke, N. and Price, C.J. (1997). “Challenges for qualitative electrical reasoning in automative
circuit simualtion”, 11th International Workshop of Qualitative Reasoning, 175- 180.
Syed Mustapha, S.M.F.D., Moseley, L.G., Jones, T.E.R., Phillips, T.N. and Price, C.J. (1998),
“Viscometric flow interpretation using qualitative and quantitative techniques”, Accepted for
publication in Engineering Applications of Artificial Intelligence.
Syed Mustapha, S.M.F.D., Phillips, T.N. and Price, C.J. (1997). “Towards characterisation of
viscoelastic behavior using model-based reasoning”, 11th International Workshop of
Qualitative Reasoning, 297-304.
Winter, H. and Baumgaertel, M. (1993). “A parsimonious model for viscoelastic liquids and
solids”, In Techniques in Rheological Measurement, chapter 5. Chapman and Hall, London.

You might also like