Professional Documents
Culture Documents
3, AUGUST 1998
Abstract— Automated model formulation is a crucial issue from using models and have made a purely experimental study
toward the construction of computational environments that can be preferred in spite of its costs and limits. Hence, the need
reason about the behavior of a physical system. The procedure arises for building intelligent computational environments that
of mathematically modeling a given physical system is quite
complex and basically involves three fundamental entities: the have the necessary methodological knowledge for both au-
experimental data, a set of candidate models, and rules for tomating the model formulation phase and assisting the user
determining in such a set the “best” model that reproduces the to properly exploit the model during his/her problem-solving
measured data. The construction of the candidate models is do- activity.
main dependent and based on specific knowledge and techniques Recent research work carried out within the qualitative
of the application domain. The choice of the best model is guided
by the data themselves; a first rough guess, which is suggested physics framework focuses on the automated model formu-
by the qualitative properties of the observed behavior, is refined lation problem and suggests two main general approaches
through system identification techniques so that the quantitative [1], [2] as well as a number of extensions and implemented
properties of the observed behavior are assessed. Therefore, systems [3]–[7]. Such an issue, which is crucial toward the
automating such a procedure requires handling and integrating
realization of tools capable to reason automatically about
different formalisms and methods, both qualitative and quantita-
tive. This paper describes a comprehensive environment that aims physical systems, involves both the construction of the model
at the automated formulation of an accurate quantitative model space for a given application domain, which we call model
of the mechanical behavior of an actual viscoelastic material library, and the selection, within such a library, of the most
in accordance with the observed response of the material to appropriate model for a given task. In the “Graph of Models”
standard experiments. To this end, algorithms and methods for
(GoM) approach [1], the model library contains complete
both the generation of an exhaustive library of models of ideal
materials and the selection of the most “accurate” model of a models of the artifact or device under study; each model, built
real material have been designed and implemented. The model manually, is characterized by different assumptions, and the
selection phase occurs in two main stages; at first, the subset of whole library is explicitly represented by a directed graph.
most plausible candidate models for the material is drawn out In this approach, model selection turns out to be a search
from the library in accordance with the qualitative properties of
the material that are highlighted by the experimental data; then,
process through the graph with the goal of finding the model
the most accurate model of the material is identified within such whose predictions are “close” to the observations [8]. In the
a set by exploiting both statistical and numerical methods. “Compositional Modeling” (CM) approach [2], the task is to
Index Terms— Automated modeling, identification, materials
formulate an appropriate model for answering a user’s query
science and technology, qualitative interpretation of data, quali- about the device behavior: such a model must be both adequate
tative simulation. and as simple as possible. The model library is organized
into pieces of knowledge, called model fragments, about
various aspects of the physical world; the model construction
I. INTRODUCTION
problem involves the selection and composition of those model
its ability both to automatically generate a library of complete A. Motivations for the Choice of the Physical Domain
models, each of them characterized by structures of different The study of materials is of great interest and practical value
complexity, and to select within it the proper one. for the variety of both methodological problems and industrial
The work described here aims at developing a comprehen- applications it offers. From a mechanical perspective, the
sive environment that automates the formulation of a model properties that characterize the responses of materials to the
of the mechanical behavior of an actual viscoelastic material action of external mechanical forces are elasticity, viscosity,
[9]. The whole formulation process occurs in the following and plasticity. As a matter of fact, the behavior of almost all
two main stages. materials is featured by the combination, at various degrees,
1) Exhaustive, but minimal, library of complete models of of the three mentioned fundamental properties. Moreover, the
ideal viscoelastic materials, which differ from each other mechanical properties of a material may be correlated to some
in structure, is automatically generated. of its other properties (for example, adhesion, hydrophilia,
2) Accurate model of an actual material is built in ac- capacity either to absorb or release active ingredients, sen-
cordance with the observed response of the material to sitivity to erosion, thyxotropy, and so on), whose knowledge
either creep or relaxation experiments. is of fundamental importance to the assessment of the material.
Each model in the library is described in two different for- These latter properties are often more difficult and expensive
malisms; at first, its structure, symbolically expressed by a to be measured than the mechanical ones. Hence, experiments
rheological formula (RF), is generated; then its correspond- that aim at measuring the mechanical features of a material are
ing parametric ordinary differential equation (ODE) is built. often preferred, as they make the assessment of the properties
The structures are automatically enumerated by analogy with of interest more economic. However, the study of a material
mechanical devices in which components that reproduce the at a purely experimental level does not allow us to capture
fundamental elastic and viscous responses are connected either the same richness of information that could be provided by
in series or in parallel. Then, by exploiting suitable connection a model. The actual difficulties in building models by hand
rules and mathematical models of the basic components, have been so far a serious deterrent from using them, in spite
the ODE model of each RF is generated. Adequate filter of their capability to make the assessment of the material
procedures, based on the algebraic properties of the con- properties more robust and the experimental protocol more
nection operators and on the mechanical equivalence that is economic. Therefore, a system that automatically maps any
captured by the ODE’s, allow us to control the combinatorial actual material to a “good” model of its mechanical behavior
explosion of the model generation process and limit the should issue a new challenge in the study of materials.
model space dimension. Moreover, we give a characterization The current implementation of the system is bounded to
of the generated ODE’s and prove [3] that the space of viscoelastic materials, that is, materials whose composite re-
possible linear models of viscoelastic materials is represented sponses are the results of a combination of elastic and viscous
by four classes of ODE’s. This is an important result since responses. Viscoelastic models describe the behavior of a wide
understanding the space of candidate models is an essential set of actual materials; for example, the mechanical behavior
step in the construction of computational environments that of polymers, whose study is of great practical importance,
aim at selecting the most appropriate model. is dominated by viscoelastic phenomena [9], [10]. The ex-
The model selection problem is solved by a mixture of tension to the plastic case, though conceptually compatible
qualitative and quantitative techniques that require both sym- with the adopted approach, still poses some major problems in
bolic and numeric computations. More precisely, we exploit representing the discontinuity of the behavior of the material.
qualitative reasoning to select the class of ODE’s that describe
the qualitative behavior of the material, and then, within the II. AUTOMATED GENERATION OF A
selected class, we identify the equation, namely, its order LIBRARY OF CANDIDATE MODELS
and the numeric values of its parameters, which refines the
quantitative properties of the material. The selection of the For the sake of clarity and completeness, we briefly recall
plausible class occurs on the ground of the comparison of the basic assumptions, definitions, and methods underlying
the simulated qualitative behaviors with the observed ones. our work. A material is assumed to be a continuous, homo-
To this end, algorithms for both qualitative simulation of the geneous, and isotropic medium, and processes are supposed
response of materials to creep and relaxation experiments, to take place in isothermal conditions so that the mechanical
and qualitative interpretation of experimental data have been aspects can be separated from the thermodynamical ones. Only
implemented. The order of the most accurate ODE model is stable materials are considered; therefore, the deformation
easily determined when the expert knows the number of either of a body solely occurs when mechanical energy has been
retardation or relaxation times [10], [11] as it can be correlated provided. Deformation phenomena are here investigated at a
with such a number. Therefore, in this case, the model selec- macroscopic level by means of macroscopic variables that
tion is reduced to a parameter identification problem. When are stress [ ] and strain [ ]. Ignoring the real internal
this information is not available, the order of the equation structure of materials has the major advantages of simplifying
is determined through an optimization technique based on the task of establishing their constitutive laws, which state a
a criterion [12] that guarantees that both the goodness of direct cause–effect relation between stress and strain. As many
experimental data fitting and the significance of the numerical rheological experiments are carried out in one dimension, that
values of the parameters themselves are balanced. is, mechanical energy is supplied through longitudinal traction
358 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART C: APPLICATIONS AND REVIEWS, VOL. 28, NO. 3, AUGUST 1998
Fig. 1. Model library generation stage: 1) RF’s are enumerated, 2) the corresponding ODE’s are generated, and 3) their qualitative behaviors are simulated.
or compression forces, we consider one-dimensional (1-D) series ( – ), each component takes the same load and
deformation processes only. the total elongation is the sum of the elongation of each
The method adopted for building the model of a material component. Summarizing, the total stress and strain can be
comes from rheology [13] and is based on a component- expressed in terms of internal variables by the following laws:
connection paradigm and internal state variables. Each fun-
damental response, in our case the elastic and viscous ones,
corresponds to an ideal material that can be represented by (parallel connection) (3)
a mechanical analogous device. More precisely, the purely
(series connection) (4)
elastic response is associated with a material analogically
represented by a spring; similarly, the purely viscous response
corresponds to a material analogically represented by a where are the internal variables, whose time
dashpot. Structures of complex materials are built by analogy evolution is expressed either directly by the basic models or
with mechanical devices, which are obtained by suitably by differential equations obtained by the recursive application
assembling, either in parallel or in series, and elements of the connection rules starting from the basic models.
so that the whole device behaves analogously to an actual The model library is automatically generated in three steps
material. The symbolic description of such an analogical and includes the following (Fig. 1):
structure, RF, represents a model of the corresponding material
1) RF’s of all unequivalent analogical structures made up
at the lowest level of description.
of components variously connected in
At a mathematical level, the pure elastic response is de-
series or in parallel;
scribed by Hooke’s law of linear elasticity
2) corresponding ODE models;
(1) 3) simulated qualitative responses of the generated models
to standard experiments.
and the viscous one by Newton’s law of linear viscosity
(2) A. Enumeration of the RF’s
where both and are positive constants that depend on The enumeration of all admissible structures is a combina-
the material and dot denotes the time derivative. Although the torial problem. Such a problem is controlled by filtering the
constitutive law of a material may be nonlinear and contain built set with respect to mechanical equivalences that are either
unconstant coefficients, we consider linear viscoelastic models, directly suggested by the algebraic properties of the connection
i.e., models whose viscoelastic behavior is described by linear operators or captured by the generated ODE’s.
ODE’s with constant coefficients. As a matter of fact, most The underlying data structures for representing a RF is
materials show a linear time-dependent behavior in the limit provided by a rooted binary tree-like graph, whose internal
of infinitesimal deformation and even in finite deformation as node labels take value in the set and leaf labels
long as the strain remains below a certain limit, which depends in the set . More precisely, we consider trees
on the material. with leaves and internal nodes that have exactly
Mathematical models of complex RF’s are built by ex- two children subtrees (Fig. 2). Let us remark that, as an
ploiting the basic models and suitable connection rules. More obvious consequence of (1)–(4), the connection operators
precisely, when components and (not necessarily basic fulfill algebraic equivalence relations, such as commutativity
components) are connected in parallel ( ), they undergo and associativity, and another property, denoted by absorption,
the same elongation, while the total stress gets distributed which states the mechanical behavioral equivalence of two
among the components; when and are connected in connected equal basic components to a single one.
CAPELO et al.: AUTOMATED FORMULATION OF ACCURATE QUANTITATIVE MODEL 359
TABLE I (5)
NUMBER OF ELEMENTS OF THE SETS Tn , Tn3 , Fn , Fn3
where and are polynomials in the time-derivative
operator, and are the variables stress and strain, respec-
tively.
Definition 2: Given the polynomials , ,
, associated with the models of the binary components
of , the polynomials are defined by
if
The enumeration process of the RF’s, or equivalently of (6)
the tree structures, occurs in four steps [3], which can be
synthesized as follows: if
1) generation of the set , where is the set of (7)
all admissible unlabeled trees that underlie -component
structures; at this stage, the attention is only focused on
the tree shape;
2) derivation of the sets and by Let us remark that Definition 2 results from the formal
filtering in accordance with a topological equivalence elimination of the internal variables by exploiting (3) and (4),
relation in , which states that unlabeled trees are respectively.
equal apart from left/right permutations; The application of the recursive algorithm defined by (6) and
3) generation of the set , where (7) and initialized by (1) and (2) to the whole set allows
is obtained by giving values to the trees us to generate the mathematical models of all ideal materials
in ; i.e., ( )-tuples of identifiers are whose structure is made up of at most components. Such
provided to label the tree nodes; ODE models, which describe the behavior of real materials
4) replacement of the set with its subset only qualitatively as long as their coefficients do not take a
, which is obtained by filtering with numeric value, can be represented by formal equations.
respect to commutativity, associativity, and absorption Definition 3: Let us call the formal equation (FE) the ODE
properties. that is obtained by giving symbolic unitary value to all nonzero
In particular, if denotes set cardinality, we have coefficients of the mathematical model.
1 As a matter of fact, the numeric values of the nonzero
and ; and are
not easily written as explicit functions of . If no filter is coefficients, which appear in the constitutive equation of
applied at Step 2), gets rapidly intractable. The algorithm the material under study, may be identified only from the
is constructive and efficient; as a matter of fact, it directly experimental data. On the other hand, at this stage of the model
builds the set by exploiting a topological characterization formulation process we do not mind knowing these values
of its elements [3]. as we focus on building an exhaustive library of candidate
Let us remark that the adopted filters dramatically reduce models within which a first selection of plausible ODE’s for
the number of the generated formulas (Table I). A further the material should be made on the ground of the qualitative
significant reduction on is subsequently obtained by a properties of the material.
filter based on the ODE structure.
C. Model Library Characterization
B. Generation of the Mathematical Models
The characterization of the space of candidate models is
For the sake of simplicity, let and still denote the essential for an efficient selection of the most accurate model.
resulting filtered sets. Given any complex formula , its Under the assumption of linear basic models, we can prove
1 Such numbers are known as Catalan numbers [11]. the following.
360 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART C: APPLICATIONS AND REVIEWS, VOL. 28, NO. 3, AUGUST 1998
Theorem 1: The set of all admissible models can be Theorem 3: The formulas , , , and
partitioned into the following four classes : map to FE , FE , FE , and FE , respec-
tively
FE FE
FE
FE FE
FE
FE The dimension of the space of mathematical models can
be derived directly from Theorems 2 and 3.
Corollary: If is the maximum number of basic compo-
FE
nents the RF’s are made of, the model space dimension is
equal to .2
where denotes the th time-derivative operator. The proof The number of equations in each class and is
of Theorem 1, which is given in [3], is derived directly from equal to if is even, and to the integer part of ,
Definition 2. otherwise; whereas, it is equal to the integer part of in
Let us remark that for , the equations FE the classes and . The index , which implicitly
model, respectively, the basic formulas , and the simplest defines the order of the ODE whose formula is made up of
composite ones . The latter two formulas are components, is equal to the integer part of or if
known in literature as Kelvin and Maxwell models, is odd or even, respectively.
respectively. Another important consequence of the previous theorems is
The formulas’ enumeration algorithm guarantees that only the following duality property.
algebraically unequivalent structures have been generated but Theorem 4: Given any arbitrarily complex formula , it
does not guarantee that all of them exhibit distinct mechanical is possible to find a mechanically equivalent formula
behaviors. Therefore, the mapping defined by the ODE obtained by combining in series a suitable number of com-
generation algorithm , where is the set of ponents , , and as well as it is possible to find an
the built mathematical models, is surjective but not injective equivalent formula obtained by combining in parallel a
and the mechanical equivalence of two different formulas can suitable number of components , , and .
be fully assessed by comparison of their respective equations. Proof:
Definition 4: Two formulas are mechanically equiv-
alent if they map to the same FE, i.e., If FE then set
. If FE then set
Next, we give theorems that state the following : 1) a one- If FE then set
to-one correspondence between the four classes of ODE’s
and suitable classes of RF’s, called reference classes, and
2) the mechanical equivalence of any complex formula with If FE then set
a formula in one of the reference classes. To this end, let us
introduce the formulas and , which are, respectively,
Let us call and equivalent dual (respectively, series
the th-order generalized and models
and parallel) representations of and still the set of all
and none but the mechanically distinct RF’s. By Theorem 4,
we can state that is isomorphic to the dual sets
Fig. 3. Standard static test: a stress or a strain is suddenly imposed and then
held constant for a time t1 = 0 t1 t0 . Fig. 4. Typical strain response to a step stress excitation (creep test).
2) Algorithm: Both and are defined by three Although in terms of computational efficiency the algorithm
logical parameters ( ) associated with either the strain is comparable to its previous version, it marks a significant
or stress properties, which take on either the value TRUE improvement: 1) it works for both creep and relaxation tests
(T) or FALSE (F). More precisely, the value T is assumed and 2) it is formally justified as it is based on sound arguments,
when the corresponding component is nonzero, and the value such as the connection rules and ODE’s. Its former version,
F, otherwise. For example, the creep response in Fig. 4 is which recursively builds the response of the material starting
qualitatively defined by (T, T, T). from the qualitative behavior of the elements and , was
In order to define and for any complex formula, partly suggested by the connection rules and partly by intuitive
remember that strains are added in series (3) and stresses are physical arguments.
added in parallel (4). Therefore Finally, let us observe that, although the algorithm has been
given, for the sake of simplicity, for input signals represented
if by step functions, it also holds for input signals represented by
then a summation of step functions. This is ensured by the linearity
of the ODE’s or, equivalently, by the Boltzmann principle of
if superposition.
then
III. FORMULATION OF AN ACCURATE
where denotes the logical OR operator. MODEL OF AN ACTUAL MATERIAL
Theorem 4 suggests a natural way of defining both In our context the model formulation process is approached
and of any given formula by considering its equivalent in the following two stages:
representation either in or in , respectively. The 1) The class of ODE models that capture the qualitative
behavior of the elements of the bases of and is derived features of the material is identified within ; let us
directly by their model, i.e., (T, F, F), denote such a class by .
(F, F, T), (F, T, F), (T, F, F), 2) is embedded in the class of ODE’s with the same
(F, F, T), and (F, T, F); whereas structure but whose coefficients are nonzero real num-
and for any complex is defined as follows: bers; the equation that refines the quantitative
properties of the material is identified within , i.e., both
its order and the numeric values of its parametric
coefficients are determined.3
Definition 5: We say that is the most accurate
model of a real material when 1) it retains the physical
qualitative properties of the material, i.e., , and
2) both the order and the numeric values of fulfill an
optimality criterion that simultaneously guarantees numerical
accuracy and parameter significance.
The algorithm strictly depends on the application domain.
However, its lack of generality is well compensated by its A. Selection of the Class of Most Plausible Candidate Models
completeness and soundness. The selection of the class of most plausible models occurs
Soundness: The proof of the soundness of the algorithm is on the ground of the qualitative comparison of the simulated
derived directly from the connection rules. As a matter of fact, behaviors with the observations (Fig. 5). Therefore, qualitative
the given simulation algorithm is the qualitative transcription interpretation of experimental data, i.e., their characterization
of the connection rules that mathematically express, through in terms of relevant qualitative physical features, is an essential
the internal variables, the links between either strain or stress step in view of model selection.
and their own respective components. With reference to the Our problem consists of defining rules for the conversion
RF’s, the strain components are added in presence of a series of the numerical observations into the three properties that
operator, whereas the stress ones in presence of a parallel characterize the qualitative profiles of the material response to
operator. the considered experiment. Quite generally, physical features
Completeness: The proof of completeness is given by The- can be captured through the identification of characteristic
orems 1–4. More precisely, Theorem 1 defines the elements shapes in the experimental data plot. Then, in order to reason
of the set of the admissible ODE models; Theorems 2 and qualitatively about the observed response, graphical data are
3, respectively, state a bijective correspondence between first abstracted to a qualitative representation; a qualitative
and (elements of both sets and ), which are built by curve description is automatically provided in terms of regions
connecting either in series the elements of the set that are homogeneous with respect to such graphical features
or in parallel the elements of the set . Finally, The- as steepness, convexity, and linearity. This process of data
orem 4 associates any given complex formula, by exploiting
its FE, with its equivalent representation both in and in . 3 Variables in bold, like p, denote vectors.
CAPELO et al.: AUTOMATED FORMULATION OF ACCURATE QUANTITATIVE MODEL 363
TABLE III As regards the creep stage, the significant time interval can
QUALITATIVE CHARACTERIZATION OF GRAPHICAL PROPERTIES be identified as follows:
FEATURED IN A TIME POINT OR INTERVAL BY MEANS
OF QUALITATIVE VALUES OF CURVE DESCRIPTORS
Fig. 8. Qualitative interpretation of the creep experimental response of a rubber-like material and plausible model selection.
scheme proposed by Klopfenstein–Reiher [21], which effec- This example also demonstrates an important attribute of the
tively works as both nonstiff and stiff ODE solver. As a matter presented approach; by limiting the candidate search space to
of fact, the ODE’s that are solved during the process may be models that are qualitatively consistent with the observations,
stiff due to the extent of elasticity featured by the modeled the modeling system is able to ensure that physical accuracy
materials. is retained throughout the quantitative refinement process.
The main problems with (9) deal with the choice of the
initial conditions; must be given to completely define the A. Model Library
problem, but the available information may be too poor to
The model library was generated by constraining the maxi-
suitably fix them. Moreover, regarding (8), a “good” guess
mum complexity of the RF’s to elements. This is by
of the solution must be provided to ensure convergence to the
far sufficient to ensure that the plausible model class, searched
real solution rather than to a local minimum. Both estimates
over during the system identification process, is numerous
for and are automatically provided by an implemented
enough to allow for structural identification.
strategy that exploits a priori knowledge and experimental
Therefore, the model space is partitioned into the following
data, as discussed below:
four classes , each one characterized by the
1) experimental curve is suitably fitted by a function creep ideal qualitative behaviors as indicated
whose shape is suggested by the abstracted qualitative
profile of the data and by the linearity assumption; (T, T, F),
2) initial condition is defined by ; FE
(T, F, F),
3) initial estimate is obtained by collocating the cur-
(F, T, T),
rent ODE on the experimental grid; i.e., is the FE
least-squares solution of the linear system obtained by (F, F, T),
substituting for in the ODE. FE (F, T, F)
(T, T, T),
FE
IV. EXAMPLE (T, F, T),
In this section, we illustrate, through a concrete example,
how the modeling tool works. Creep experimental data, re- B. Plausible Model Selection
lated to a rubber-like material, were kindly provided by the The experimental curve and the computer outcome of its
Scientific Division of WPABISCH S.p.A., Italy. The creep qualitative interpretation are shown in Fig. 8.
experiment was carried out at 50 C, over a time range of 141 Creep and recovery graphical data are analyzed, and a qual-
s, with a constant stress value of 200 Pa. itative description is produced in terms of a list of qualitative
CAPELO et al.: AUTOMATED FORMULATION OF ACCURATE QUANTITATIVE MODEL 367
0
Fig. 9. Measured ( ) versus predicted ( ) strain, according to the most accurate model identified by the system.
qualitative reasoning for both an efficient and physically formulation task is reduced to a search problem through the
correct approach, though the goal is the formulation of an graph and aims at finding the model whose predicted behavior
accurate quantitative model. An automated search, performed is “close” to the observed one. The procedure starts from
“blindly” over the whole library at a purely numeric level, the simplest model, and a new one is searched when the
may yield a model that best fits the observations but does not model predictions entailed by the initial choice do not match
capture all of the qualitative features of the physical system the desired accuracy versus the observations. The selection
at study, such as, for example, discontinuities of the behavior. of a new model is driven by domain-dependent “parameter-
The search model set is properly chosen within the model change” rules, which specify the qualitative effects of each
space for the domain when it satisfies a qualitative accuracy assumption on the model’s predictions, and necessarily re-
criterion; i.e., any of its elements qualitatively represents quires that all possible consistent combinations of modeling
the physical properties captured by the observations. Besides assumptions are explicitly stored. In the GoM framework,
the evident advantage that derives from this first selection, both the models and the parameter-change rules associated
the computational costs are significantly reduced because with assumption transitions must be provided by hand by the
of the reduced dimension of the search space. Moreover, model builder. This heavy modeling work together with the
the restriction of the search space to all and none but the storage requirement, which is exponential in the number of
qualitatively meaningful models allows us to better delimit assumptions, may render problematic the actual applicability
the a priori knowledge that could be conveniently exploited of such a conceptually interesting approach. On the other hand,
in the next steps of the automated modeling process. More providing the GoM approach with the ability to automate the
precisely, such knowledge may suggest a proper choice of modeling step appears unfeasible. The main reason, due to its
either the initial conditions needed for solving the initial value generality, deals with the reformulation of a new model when
problem or a good initial guess for the model parameters. moving from one node to another; though the two models share
The qualitative reasoning methods discussed here deal with most variables, they may be represented by quite different
both simulation and data interpretation. The simulation al- equations and then require, for their reformulation, knowledge
gorithm is strictly domain dependent but generates all and about the physics of the system at study as well as complex
none but the actual physical behaviors. The data interpretation problem solving techniques. A domain-independent technique
algorithm provides reasoning techniques to emulate the expert- focused on the single aspect of automatically shifting model
like visual interpretation of experimental data. In spite of accuracy has been proposed in [8]. The core of such a
its simplicity, the algorithm also provides useful information technique, which provides implemented algorithms for select-
about the adequacy of the models in the library to describe ing, evaluating, and switching modeling assumptions, deals
the behavior of the material under study; a possible convex with the automated generation of the qualitative behavioral
shape of the data during the creep phase denotes that the differences produced by two models, information that is given
linear theory of elasticity and viscosity we adopted to build in [1] by the hand-crafted parameter-change rules. Notice
the model library is inadequate to study such a material. that, in these papers, the term accuracy does not refer to the
Although it has been designed for interpreting data in a specific resolution of model accuracy, i.e., the numeric precision of
domain, it becomes a domain-independent technique, as it the model predictions versus the observations, but rather to
is capable of transforming a stream of observed data into a the adequacy of the model to represent the real system.
qualitative description that characterizes its shape, that is, it As in the GoM framework, our approach relies on a pre-
highlights the qualitative properties of the numerical data, such defined library of complete models and aims at searching the
as monotonicity, convexity, and linearity regions. library for the model that best fits the observations. But, unlike
GoM, all models are automatically generated from primitive
models of the considered domain and the quantitative model
VI. RELATED WORK we look for is required to be accurate, in regard to both
Several pieces of work have addressed the model formula- physical adequacy and numerical resolution.
tion problem. Most of them focus on one or more aspects of In our work the lack of generality turns out to be very well
reasoning with multiple models, and aim at providing either compensated by the availability of specific domain knowledge.
the GoM [1] or the CM [2] approach, which can be rightly The exploitation of such knowledge has made it possible
considered seminal works for automated modeling, with an to characterize the model library as well as to manage and
efficient model selection strategy. Our method lies halfway organize the generated models so that, although the model
between the GoM and CM approaches; it is closer to GoM, in enumeration process in itself has exponential complexity, the
regard to the model selection goal, whereas it is similar to CM, space required for their storage is linear in the number of
in regard to the model construction from primitive models. primitive components. As a consequence, the search process
Its direct comparison with both of them should highlight the has linear complexity. The model selection and the system
differences. identification tasks take advantage, respectively, of the parti-
In the GoM framework, the library of possible models is tion of the model library into classes of models featured by
represented by a directed graph, where each node represents the same qualitative behavior and of the hierarchical structure
a complete model of the system at study and each edge of the plausible model class.
connecting two nodes identifies the assumption (i.e., the ap- In the CM approach, the domain knowledge is decomposed
proximation) that distinguishes the two models. The model by using explicit assumptions into pieces of knowledge, called
CAPELO et al.: AUTOMATED FORMULATION OF ACCURATE QUANTITATIVE MODEL 369
model fragments, each of them describing a conceptually once and for all, rather than building the constitutive equation
distinct physical object or phenomenon. The organization of from scratch whenever a new material is analyzed, makes the
a domain theory into model fragments allows their reusability use of our modeling tool time effective within an industrial
for a large space of different scenarios. A model fragment application context. As a matter of fact, to provide a tool that
consists of conditions and consequences, which, respectively, automatically generates an accurate model from data is the
specify the conditions under which a phenomenon occurs and added value given by our work to the analysis of a material
the functional relations between the attributes of the objects for its industrial assessment.
involved in the phenomenon. The formulation problem in this
framework is addressed as follows: given a description of an VII. CONCLUSION
artifact or, more generally, of the structure of a physical system
and a specific query about some aspects of its behavior, gen- This paper discusses a variety of qualitative, quantitative,
and model-based reasoning techniques to automate the whole
erate a model to answer the query by selecting and composing
process of accurate mathematical modeling from empirical
appropriate model fragments. Therefore, such a model is not
data. Such techniques, ad hoc developed for an application
explicitly predefined in a library, but it is formulated from
domain that actually presents difficulties in building models by
scratch in response to a user’s query. Two different issues must
hand, concern model library generation, qualitative analysis of
be considered in the compositional process: the constructed
empirical data, simulation to generate qualitative profiles of the
model has to be both adequate and as simple as possible. To
solutions of model equations, and identification of the model
outline, the compositional algorithm consists of four steps.
that most accurately explains the observed behavior.
The first step identifies, through a syntactic analysis of the
The integration of different formalisms and methods has
query, the set of objects and terms of interest. Then, the object
allowed us to emulate the expert’s way of reasoning about
expansion step identifies, by using the part-of hierarchy, all
materials from a rheological perspective with the goal of
additional components of the system that should be included
associating a real material with its constitutive law. A natural
in the initial set of objects to capture all relevant interactions.
use of such an equation deals with the simulation of the
In the third step, all internally consistent and complete sets
behavior of the material under complex loads, but more
of candidate model fragments are generated. Finally, one
importantly, it can be exploited for a precise analysis of
set is chosen according to simplicity criteria. Notice that
the material during its assessment phase; the parameters in
such a procedure is time exponential. Moreover, CM does
the equation could be interpreted as a measure of material
not guarantee the inclusion of all and none but the model
properties other than the rheological ones correlated with
fragments causally related to the query and the sufficiency of
them. However, the latter use of the model requires the
the model. Several pieces of work have contributed to improve
definition of a rationale that allows us to correlate the model
the efficiency of CM, as far as both the additional knowledge
parameter vector and physical properties of materials. Besides
[22] and the selection of relevant model fragments [23], [24]
a new challenge to the study of materials in their design
are concerned.
and assessment phase, our work also provides a small but
Similarly to CM, our approach builds a model by compos-
significant contribution to the application domain itself; the
ing primitive models of the domain. Such primitive models
generation, and the consequent analysis, of models of high-
describe the behavior of the single components of a device
structural complexity has allowed us to discover that the whole
that, as a whole, behaves analogously to the material. In
set of candidate linear models is mathematically represented
this sense, the device provides a structural description of the
by four classes of ODE’s, each of them describes a class
material whose real internal structure is unknown. Modeling by
of viscoelastic materials of different structural complexity
“analogous device” is a common practice in classical modeling
but characterized by the same qualitative behavior. Such a
frameworks whenever the internal structure of the system
result is not explicitly reported in the literature even if it
is unknown. For example, a well-known technique in the
is likely inferable from it. As far as its practical use is
biomedical domain is compartmental modeling [5], in which
concerned, the current version of our system is applicable to a
a system is described as a finite set of interacting components
wide set of actual materials, for example, polymers, whose
that are generally idealized stores of substances.
behavior is dominated by viscoelastic phenomena, and its
Remember that our task is to build a quantitative accurate
performance has been demonstrated by a number of tests on
model of an actual material given only a set of its observations
data collected through creep experiments on materials such as
in response to standard experiments. Then, building a model
inks, rubbers, and drugs. Nevertheless, the extension of the
means to also identify its structure. If we did not care about
model library to both nonlinear models, which can be built
numerical resolution, as in CM, a generation from scratch
starting from nonlinear laws, e.g.,
would provide, exponentially in time, for the structurally
and for the elastic and viscous basic
simplest model physically coherent with the observations. On
components, respectively, and to models that also take plas-
the other hand, we cannot give up numerical resolution, as it
ticity into account will make the system a more general tool.
ensures that the structural complexity of the material, which
gives information about some of its physical properties, is
captured. The generation from scratch to also meet numerical ACKNOWLEDGMENT
accuracy is feasible, in principle, but not advisable, in practice, The authors would especially like to thank C. Caramella of
due to its prohibitive cost. The generation of the model library the Department of Pharmaceutical Chemistry of the University
370 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART C: APPLICATIONS AND REVIEWS, VOL. 28, NO. 3, AUGUST 1998
of Pavia and B. Pirotti of the WPABISCH S.p.A—Scientific [21] R. W. Klopfenstein, “Numerical differentiation formulas for stiff sys-
Division, who provided the experimental data. They are also tems of ordinary differential equations,” RCA Rev., vol. 32, pp. 447–462,
1971.
grateful to P. C. Franzone for helpful discussions and useful [22] P. P. Nayak, “Causal approximations,” Artif. Intell., vol. 70, pp.
comments. 277–334, 1994.
[23] Y. Iwasaki and A. Y. Levy, “Automated model selection for simulation,”
in Proc. AAAI-94. Los Altos, CA: Morgan Kaufmann.
REFERENCES [24] J. Rickel and B. Porter, “Automated modeling for answering prediction
questions: Selecting the time scale and system boundary,” in Proc.
[1] S. R. Addanki, R. Cremonini, and J. S. Penberthy, “Graphs of models,” AAAI-94. Los Altos, CA: Morgan Kaufmann.
Artif. Intell., vol. 51, pp. 145–177, 1991.
[2] B. Falkenhaier and K. D. Forbus, “Compositional modeling: Finding the
right model for the job,” Artif. Intell., vol. 51, pp. 95–143, 1991.
[3] A. C. Capelo, L. Ironi, and S. Tentoni, “A model-based system for the Antonio C. Capelo received the degree in electrical
classification and analysis of materials,” Intell. Syst. Eng., vol. 2, no. 3, engineering from the Technical University of Lis-
pp. 145–158, 1993. bon, Lisbon, Portugal, in 1972.
[4] J. Crawford, A. Farquhar, and B. J. Kuipers, “QPC: A compiler He is Professor of Mathematical Analysis at the
from physical models into qualitative differential equations,” in Re- Faculty of Statistics, University of Padua, Padua,
cent Advances in Qualitative Physics, P. Struss and B. Faltings, Eds. Italy. He has had a research collaboration contract
Cambridge, MA: MIT Press, 1992, pp. 17–32. with the Institute of Numerical Analysis, National
[5] L. Ironi and M. Stefanelli, “A framework for building and simulating Council of Researches (IAN-CNR), Pavia, Italy,
qualitative models of compartmental systems,” Computat. Methods since 1994. His major fields of interest are math-
Programs Biomed., vol. 42, pp. 233–254, 1994. ematical and statistical modeling in engineering,
[6] Y. Iwasaki, “Reasoning with multiple abstraction models,” in Recent
biomedical sciences, and environment. He is the
Advances in Qualitative Physics, B. Faltings and P. Struss, Eds. Cam-
author of books and papers on pure and applied mathematics.
bridge, MA: MIT Press, 1992, pp. 67–82.
[7] C. M. Low and Y. Iwasaki, “Device modeling environment: An interac-
tive environment for modeling device behavior,” Intell. Syst. Eng., vol.
1, no. 2, pp. 115–145, 1992.
[8] D. S. Weld, “Reasoning about model accuracy,” Artif. Intell., vol. 56, Liliana Ironi received the degree (Laurea) in math-
pp. 255–300, 1992. ematics from the University of Pavia, Pavia, Italy,
[9] R. G. C. Arridge, Mechanics of Polymers. Oxford, U.K.: Clarendon, in 1972.
1975. She is currently a Senior Research Scientist at the
[10] J. D. Ferry, Viscoelastic Properties of Polymers. New York: Wiley, Institute of Numerical Analysis, National Council of
1970. Researches (IAN-CNR), Pavia, and coordinates the
[11] R. Whorlow, Rheological Techniques. Chichester, U.K.: Ellis Hor- research project on “Methods and Algorithms for
wood, 1980. Qualitative Simulation” at IAN. Her research work
[12] H. Akaike, “A new look at the statistical model identification,” IEEE
includes mathematical modeling, both quantitative
Trans. Automat. Contr., vol. AC-19, pp. 716–723, June 1974.
and qualitative, for different domains, such as engi-
[13] M. Reiner, “Rheology,” in Encyclopedia of Physics, S. Flügge, Ed.
neering, biology, and medicine. She has published
Berlin, Germany: Springer-Verlag, 1958, vol. 6, pp. 434–550.
[14] B. J. Kuipers, “Qualitative simulation,” Artif. Intell., vol. 29, pp. extensively on topics within applied mathematics and artificial intelligence.
280–338, 1986. Dr. Ironi has served on the program committees of several international
[15] K. D. Forbus, “Interpreting observations of physical systems,” IEEE conferences.
Trans. Syst., Man, Cybern., vol. SMC-17, pp. 350–359, Mar. 1987.
[16] D. DeCoste, “Dynamic across-time measurement interpretation,” Artif.
Intell., vol. 51, pp. 273–342, 1991.
[17] S. A. McIlraith, “Qualitative data modeling: Application of a mechanism Stefania Tentoni received the degree in mathemat-
for interpreting graphical data,” Computat. Intell., vol. 5, pp. 111–120, ics from the University of Pavia, Pavia, Italy, in
1989. 1980.
[18] L. Ljung, System Identification. Englewood Cliffs, NJ: Prentice-Hall, She is currently a Research Scientist at the In-
1987. stitute of Numerical Analysis, National Council of
[19] B. Choi, ARMA Model Identification. New York: Springer-Verlag, Researches (IAN-CNR), Pavia. Her main research
1992. interests are in the field of mathematical modeling,
[20] J. J. Moré, “The Levenberg–Marquardt algorithm: Implementation and with a focus on the integration of qualitative and
theory,” in Numerical Analysis, Lecture Notes in Mathematics, vol. quantitative techniques.
630, G. A. Watson, Ed. Berlin, Germany: Springer-Verlag, 1977, pp.
105–116.