GIREP INTERNATIONAL COPNFERENCE

Modelling in Physics and Physics Education

20.-25. August 2006, AMSTEL institute, Faculty of Science, University of Amsterdam, Netherlands

Notes for a Modeling Theory of Science, Cognition and

Instruction

David Hestenes
Arizona State University

Abstract
Modeling Theory provides common ground for interdisciplinary research in
science education and the many branches of cognitive science, with
implications for scientific practice, instructional design, and connections
between science, mathematics and common sense.

I. Introduction

During the last two decades Physics

Education Research (PER) has To the grand philosophical question: “What
emerged as a viable sub-discipline is man?”
Aristotle answered:
of physics, with faculty in physics “Man is a rational animal.”
departments specializing in research Modeling Theory offers a new answer:
on learning and teaching physics. “Man is a modeling animal!”
There is still plenty of resistance to Homo modelus!
PER from hard-nosed physicists
who are suspicious of any research that smacks of education, psychology or
philosophy. However, that is countered by a growing body of results
documenting deficiencies in traditional physics instruction and significant
improvements with PER-based pedagogy. Overall, PER supports the general
conclusion that science content cannot be separated from pedagogy in the
design of effective science instruction. Student learning depends as much on
structure and organization of subject matter as on the mode of student
engagement. For this reason, science education research must be located in
science departments and not consigned to colleges of education.
As one of the players in PER from its beginning, my main concern has
been to establish a scientific theory of instruction to guide research and practice.
Drawing on my own experience as a research scientist, I identified construction
and use of conceptual models as central to scientific research and practice, so I
adopted it as the thematic core for a MODELING THEORY of science
instruction. From the beginning, it was clear that Modeling Theory had to
address cognition and learning in everyday life as well as in science, so it
required development of a model-based epistemology and philosophy of
science. Thus began a theory-driven MODELING RESEARCH PROGRAM:
Applying the theory to design curriculum and instruction, evaluating results,
and revising theory and teaching methods accordingly. Fig. 1 provides an
overview of the program.

34
 Section II reviews evolution of the Modeling Research Program.
Concurrent evolution of Cognitive Science is outlined in Section III. Then
comes the main purpose of this paper: To lay foundations for a common
modeling theory in cognitive science and science education to drive symbiotic
research in both fields. Specific research in both fields is then directed toward a
unified account of cognition in common sense, science and mathematics. This
opens enormous opportunities for science education research that I hope some
readers will be induced to pursue.
Of course, I am not alone in recognizing the importance of models and
modeling in science, cognition, and instruction. Since this theme cuts across the
whole of science, I have surely overlooked many important insights. I can only
hope that this paper contributes to a broader dialog if not to common research
objectives.

Modeling Theory: Research & Practice

Scientific Research

Cognitive science Scientific practice

Modeling Theory Modeling Theory

of cognition and learning of scientific knowledge

Instructional Design

Fig. 1: Teaching Practice

II. Evolution of Modeling Instruction

My abiding interest in questions about cognition and epistemology in science

and mathematics was initiated by undergraduate studies in philosophy. In 1956 I
switched to graduate studies in physics with the hope of finding some answers.
By 1976 I had established a productive research program in theoretical physics
and mathematics, which, I am pleased to say, is still flourishing today [1].
About that time, activities of my colleagues Richard Stoner, Bill Tillery and
Anton Lawson provoked my interest in problems of student learning. The result
was my first article advocating Physics Education Research [2].
I was soon forced to follow my own advice by the responsibility of
directing PER doctoral dissertations for two outstanding graduate students,

35
Ibrahim Halloun and Malcolm Wells. Halloun started about a year before Wells.
In my interaction with them, two major research themes emerged: First, effects
on student learning of organizing instruction about models and modeling;
Second, effects of instruction on student preconceptions about physics.
The Modeling Instruction theme came easy. I was already convinced of
the central role of modeling in physics research, and I had nearly completed an
advanced monograph-textbook on classical mechanics with a modeling
emphasis [12]. So, with Halloun as helpful teaching assistant, I conducted
several years of experiments with modeling in my introductory physics courses.
The second theme was more problematic. I was led to focus on modes of
student thinking by numerous discussions with Richard Stoner about results
from exams in his introductory physics course. His exam questions called for
qualitative answers only, because he believed that is a better indicator of physics
understanding than quantitative problem solving. However, despite his heroic
efforts to improve every aspect of his course, from the design of labs and
problem solving activities to personal interaction with students, class average
scores on his exams remained consistently below 40%. In our lengthy
discussions of student responses to his questions, I was struck by what they
revealed about student thinking and its divergence from the physics he was
trying to teach. So I resolved to design a test to evaluate the discrepancy
systematically. During the next several years I encountered numerous hints in
the literature on what to include. When Halloun arrived, I turned the project
over to him to complete the hard work of designing test items, validating the
test and analyzing test results from a large body of students.
The results [3, 4] were a stunning surprise! surprising even me! so
stunning that the journal editor accelerated publication! With subsequent
improvements [5], the test is now known as the Force Concept Inventory (FCI),
but that has only consolidated and enhanced the initial results. Instructional
implications are discussed below in connection with recent developments. For
the moment, it suffices to know that the FCI was immediately recognized as a
reliable instrument for evaluating the effectiveness of introductory physics
instruction in both high school and college.
Five major papers [6-10] have been published on Modeling Theory and
its application to instruction. These papers provide the theoretical backbone for
the Modeling Instruction Project [11], which is arguably the most successful
program for high school physics reform in the U.S. if not the world. Since the
papers have been seldom noted outside that project, a few words about what
they offer is in order.
The first paper [6] provides the initial theoretical foundation for Modeling
Theory and its relation to cognitive science. As modeling has become a popular
theme in science education in recent years, it may be hard to understand the
resistance it met in 1985 when my paper was first submitted. Publication was
delayed for two years by vehement objections of a referee who was finally
overruled by the editor. Subsequently, the paper was dismissed as mere
speculation by empiricists in the PER community, despite the fact that it was
accompanied by a paper documenting successful application to instruction.

36
Nevertheless, this paper provided the initial conceptual framework for all
subsequent developments in modeling instruction. It must be admitted, though,
the paper is a difficult read, more appropriate for researchers than teachers.
Paper [7] is my personal favorite in the lot, because it exorcises the
accumulated positivist contamination of Newtonian physics in favor of a model-
centered cognitive account. For the first time it breaks with tradition to
formulate all six of Newton’s laws. This is important pedagogically, because all
six laws were needed for complete coverage of the “Force concept” in designing
the FCI [5]. Moreover, explicit formulation of the Zeroth Law (about space and
time) should interest all physicists, because that is the part of Newtonian
physics that was changed by relativity theory. Beyond that, the paper shows that
Newton consciously employed basic modeling techniques with great skill and
insight. Indeed, Newton can be credited with formulating the first set of rules
for MODELING GAMES that scientists have been playing ever since.
Paper [11] applies Modeling Theory to instructional design, especially the
design of software to facilitate modeling activities. Unfortunately, the R&D
necessary to build such software is very expensive, and funding sources are still
not geared to support it.
In contrast to the preceding theoretical emphasis, papers [8, 9] are aimed
at practicing teachers. Paper [8] describes the results of Wells’ doctoral thesis,
along with instructional design that he and I worked out together and his
brilliant innovations in modeling discourse management. His invention of the
portable whiteboard to organize student discourse is propagating to classrooms
throughout the world. Sadly, terminal illness prevented him from contributing to
this account of his work.
Wells’ doctoral research deserves recognition as one of the most
successful and significant pedagogical experiments ever conducted. He came to
me as an accomplished teacher with 30 years experience who had explored
every available teaching resource. He had already created a complete system of
activities to support student-centered inquiry that fulfills every recommendation
of the National Science Education Standards today. Still he was unsatisfied.
Stunned by the performance of his students on the FCI-precursor, he resolved to
adapt to high school the ideas of modeling instruction that Halloun and I were
experimenting with in college. The controls for his experiment were
exceptional. As one control, he had complete data on performance of his own
students without modeling. Classroom activities for treatment and control
groups were identical. The only difference was that discourse and activities
were focused on models with emphasis on eliciting and evaluating the students’
own ideas. As a second control, posttest results for the treatment group were
compared to a well-matched group taught by traditional methods over the same
time period. The comparative performance gains of his students were
unprecedented. However, I am absolutely confident of their validity, because
they have been duplicated many times, not only by Wells but others that
followed.
I was so impressed with Wells’ results that I obtained in 1989 a grant
from the U.S. National Science Foundation, to help him develop Modeling

37
Workshops to inspire and enable other teachers to duplicate his feat. Thus began
the Modeling Instruction Project, which, with continuous NSF support, has
evolved through several stages with progressively broader implications for
science education reform throughout the United States. Details are available at
the project website [11]. None of this, including my own involvement, would
have happened without the pioneering influence of Malcolm Wells.

III. Evolution of Cognitive Science

Cognitive science grew up in parallel with PER and Modeling Theory. With the
aim of connecting the strands, let me describe the emergence of cognitive
science from the perspective of one who has followed these developments from
the beginning. Of course, the mysteries of the human thought have been the
subject of philosophical contemplation since ancient times, but sufficient
empirical and theoretical resources to support a genuine science of mind have
been assembled only recently. Box 1 outlines the main points I want to make.
I regard the formalist movement in mathematics as an essential
component in the evolution of mathematics as the science of structure, which is
a central theme in our formulation of Modeling Theory below. Axioms are often
dismissed as mathematical niceties, inessential to science. But it should be
recognized that axioms are essential to Euclidean geometry, and without
geometry there is no science. I believe that the central figure in the formalist
movement, David Hilbert, was the first to recognize that axioms are actually
definitions! Axioms define the structure in a mathematical system, and
structure makes rational inference possible!
Equally important to science is the operational structure of scientific
measurement, for this is essential to relate theoretical structures to experiential
structures in the physical world. This point has been made most emphatically by
physicist Percy Bridgeman, with his concept of operational definitions for
physical quantities (but see [7] for qualifications). However, to my mind, the
deepest analysis of scientific measurement has been made by Henri Poincaré,
who explained how measurement conventions profoundly influence theoretical
conceptions. In particular, he claimed that curvature of physical space is not a
fact of nature independent of how measurements are defined. This claim has
long been inconclusively debated in philosophical circles, but recently it
received spectacular confirmation [14].
Following a long tradition in rationalist philosophy, the formalist
movement in mathematics and logic has been widely construed as the
foundation for a theory of mind, especially in Anglo-American analytic
philosophy. This is an egregious mistake that has been roundly criticized by
George Lakoff and Mark Johnson [17-21] in the light of recent developments in
cognitive science. Even so, as already suggested, formalist notions play an
important role in characterizing structure in cognition.
The creation of serial computers can be construed as technological
implementation of operational structures developed in the formalist tradition. It

38
soon stimulated the
creation of information Box 1: Emergence of Cognitive Science
processing psychology, I. Scientific Precursors
with the notion that • Formalist mathematics and logic (~1850-1940)
– axioms & standards for rigorous proof
cognition is all about – reasoning by rules and algorithms
symbol processing. I was • Operationalism (Bridgeman, 1930)
right up to date in • Conventionalism (Poincaré, 1902)
applying this egregious • Gestalt psychology (~1915-1940)
mistake to physics • Genetic Epistemology (Piaget, ~1930-1960)
teaching [2]. Even so, II. Emergence of computers and computer science
(~1945-1970) implementing operational structures
most of the important III. First Generation Cognitive Science (~1960-1980)
research results and • “Brain is a serial computer” metaphor
insights that I reported • “Mind is a computer software system”
survive reinterpretation • Information processing psychology & AI
when the confusion – Thinking is symbol manipulation
between cognition and – functionalism (details about the brain irrelevant)
IV. Second Generation Cognitive Science (~1983-- )
symbol processing is • Neural network level
straightened out. Symbol – Brain is a massively parallel dynamical system
processing is still a – Thinking is pattern processing
central idea in computer • Cognitive phenomenology at the functional level:
science and Artificial – empirical evidence for mental modeling
Intelligence (AI), but is accumulating rapidly from many sources.
only the ill-informed
confuse it with cognitive processes.
I tried to link the dates in Box 1 to significant events in each category. I
selected the date 1983 for the onset of second generation cognitive science,
because I had the privilege of co-organizing the very first conference devoted
exclusively to what is now known as cognitive neuroscience. It still took several
years to overcome the heavy empiricist bias of the neuroscience community and
establish neural network modeling as a respectable activity in the field. The
consequence has been a revolution in thinking about thinking that we aim to
exploit.
IV. Modeling Research in Cognitive Science
With its promise for a universal science of mind, research in cognitive science
cuts across every scientific discipline and beyond. Box 2 lists research that I see
as highly relevant to the Modeling Theory I am promoting. The list is
illustrative only, as many of my favorites are omitted. These scientists are so
productive that it is impractical to cite even their most important work. Instead,
I call attention to the various research themes, which will be expanded with
citations when specifics are discussed.
References [15, 16] provide an entrée to the important work of Giere,
Nercessian and Gentner, which has so much in common with my own thinking
that it may be hard to believe it developed independently. This illustrates the
fact that significant ideas are implicit in the culture of science waiting for
investigators to explicate and cultivate as their own.

39
 In sections to follow, I emphasize alignment of Modeling Theory with
Cognitive Linguistics, especially as expounded by George Lakoff [17-20].
Language is a window to the mind, and linguistic research has distilled a vast
corpus of data to deep insights into structure and use of language. My objective
is to apply these insights to understanding cognition in science and
mathematics. Cognitive Linguistics makes this possible, because it is a
reconstruction of linguistic theory aligned with the recent revolution in
Cognitive Science.
V. Constraints from Cognitive Neuroscience
Box 2: Modeling Research in Cognitive Science
Cognitive neuroscience is
concerned with explaining Philosophy of Science
cognition as a function of the Ronald Giere (Model-based philosophy of science)
brain. It bridges the interface Jon Barwise (Deductive inference from diagrams)
between psychology and
History and Sociology of Science
biology. The problem is to Thomas Kuhn (Research driven by Exemplars)
match cognitive theory at the Nancy Nercessian (Maxwell’s analogical modeling)
psychological level with
neural network theory at the Cognitive Psychology
biological level. Already there Dedre Gentner (Analogical reasoning)
Philip Johnson-Laird (Inference with mental
is considerable evidence
models)
supporting the working Barbara Tversky (Spatial mental models vs. visual
hypothesis that cognition (at imagery)
the psychological level) is
grounded in the sensory- Cognitive Linguistics
motor system (at the George Lakoff (Metaphors & radial categories)
Ronald Langacker (Cognitive grammar & image
biological level).
schemas)
The evidence is of three
kinds: Cognitive Neuroscience
x Soft constraints: Michael O’Keeffe (Hippocampus as a Cognitive
Validated models of Map)
cognitive structure Stephen Grossberg (Neural network theory)
from cognitive science, Physics Education Research
especially cognitive Andy diSessa (Phenomenological primitives)
linguistics. John Clement (Bridging analogies)
x Hard constraints: Information & Design Sciences
Identification of UML: Universal Modeling Language
& Object-Oriented Programming
specific neural
architectures and
mechanisms sufficient to support cognition and memory.
x Evolutionary constraint: A plausible account of how the brain could have
evolved to support cognition.
A few comments will help fix some of the issues.
Biology tells us that brains evolved adaptively to enable navigation to
find food and respond to threats. Perception and action are surely grounded in

40
identifiable brain structures of the sensory-motor system. However, no
comparable brain structures specialized for cognition have been identified. This
strongly suggests that cognition too is grounded in the sensory-motor system.
The main question is then: what adaptations and extensions of the sensory-
motor system are necessary to support cognition?
I hold that introspection, despite its bad scientific reputation, is a crucial
source of information about cognition that has been systematically explored
by philosophers, linguists and mathematicians for ages. As Kant was first to
realize and Lakoff has recently elaborated [20], the very structure of
mathematics is shaped by hard constraints on the way we think. A major
conclusion is that geometric concepts (grounded in the sensory-motor
system) are the prime source of relational structures in mathematical
systems.
I am in general agreement with Mark Johnson’s NeoKantian account of
cognition [21], which draws on soft constraints from Cognitive Linguistics. But
it needs support by reconciliation with hard constraints from sensory-motor
neuroscience. That defines a promising direction for research in Cognitive
Neuroscience. Let me reiterate my firm opinion [6] that the research program of
Stephen Grossberg provides the best theoretical resources to pursue it.
VI. System, Model & Theory; Structure & Morphism

The terms ‘system’ and

Box 3: A conceptual MODEL is defined by specifying
‘model’ have been five types of structure:
ubiquitous in science
and engineering since (a) systemic structure:
the middle of the x composition (internal parts (objects) in the
twentieth century. system)
Mostly these terms are x environment (external agents linked to the
system)
used informally, so
x connections (external and internal links)
their meanings are (b) geometric structure:
quite variable. But for x position with respect to a reference frame
the purposes of (external)
Modeling Theory, we x configuration (geometric relations among the
need to define them as parts)
sharply as possible. (c) object structure:
Without duplicating my x intrinsic properties of the parts
(d) interaction structure:
lengthy discussions of
x properties of (causal) links
this matter before [7- (e) temporal (event) structure:
10], let me reiterate x temporal change in structure of the system
some key points with
an eye to preparing a deeper connection to cognitive theory in the next section.
I define a SYSTEM as a set of related objects. Systems can be of any
kind depending on the kind of object. A system itself is an object, and the
objects of which it is composed may be systems. In a conceptual system the
objects are concepts. In a material system the objects are material things. Unless

41
otherwise indicated, we assume that the systems we are talking about are
material systems. A material system can be classified as physical, chemical or
biological, depending on relations and properties attributed to the objects.
The STRUCTURE of a system is defined as the set of relations among
objects in the system. This includes the relation of “belonging to,” which
specifies COMPOSITION, the set of objects belonging to the system. A
universal finding of science is that all material systems have geometric, causal
and temporal structure, and no other (metaphysical) properties are needed to
account for their behavior. According to Modeling Theory, science comes to
know objects in the real world not by direct observation, but by constructing
conceptual models to interpret observations and represent the objects in the
mind. This epistemological precept is called Constructive Realism by
philosopher Ronald Giere.
I define a conceptual MODEL as a representation of structure in a
material system, which may be real or imaginary. The possible types of
structure are summarized in Box 3. I have been using this definition of model
for a long time, and I am yet to find a model in any branch of science that
cannot be expressed in these terms.
Models are of many kinds, depending on their purpose. All models are
idealizations, representing only structure that is relevant to the purpose, not
necessarily including all five types of structure in Box 3. The prototypical kind
of model is a map. Its main purpose is to specify geometric structure (relations
among places), though it also specifies objects in various locations. Maps can be
extended to represent motion of an object by a path on the map. I call such a
model a motion map. Motion maps should not be confused with graphs of
motion, though this point is seldom made in physics or math courses. In
relativity theory, motion maps and graphs are combined in a single spacetime
map to represent integrated spatiotemporal event structure.
A mathematical model represents the structure of a system by
quantitative variables of two types: state variables, specifying composition,
geometry and object properties; interaction variables, specifying links among
the parts and with the environment [6]. A process model represents temporal
structure as change of state variables. There are two types. A descriptive model
represents change by explicit functions of time. A dynamical model specifies
equations of change determined by interaction laws. Interaction laws express
interaction variables as functions of state variables.
A scientific THEORY is defined by a system of general principles (or
Laws) specifying a class of state variables, interactions and dynamics (modes of
change) [6, 7]. Scientific practice is governed by two kinds of law:
I. Statutes: General Laws defining the domain and structure of a Theory
(such as Newton’s Laws and Maxwell’s equations)
II. Ordinances: Specific laws defining models
(such as Galileo’s law of falling bodies and Snell’s law)

42
The content of a scientific theory is a population of validated models. The
statutes of a theory can be validated only indirectly through validation of
models.
Laws defining state variables are intimately related to Principles of
Measurement (also called correspondence rules or operational definitions) for
assigning measured values to states of a system. A model is validated to the
degree that measured values (data) match predicted values determined by the
model. The class of systems and range of variables that match a given model is
called its domain of validity. The domain of validity for a theory is the union of
the validity domains for its models.
Conceptual
Conceptual Analogy
World: Model I Model II

Referential
Material Analogy
Material Analogy
World: System I System II

Fig. 3: Three Kinds of Analogy

Empirical observation and measurement determine an analogy between a

given model and its referent (a system). I call this a referential analogy. An
analogy is defined as a mapping of structure from one domain (source) to
another (target). The mapping is always partial, which means that some
structure is not mapped. (For alternative views on analogy see [16].) Analogy is
ubiquitous in science, but often goes unnoticed. Several different kinds are
illustrated in Figs. 3&4.

Conceptual analogies
between models in different Model I
domains are common in science
and often play a generative role Referential
in research. Maxwell, for Analogy Inductive
example, explicitly exploited Analogy
System I System II
electrical– mechanical analogies.
An analogy specifies differences Fig. 4: Material equivalence
as well as similarities between
source and target. For example,
similar models of wave propagation for light, sound and water and ropes
suppress confounding differences, such as the role of an underlying medium.
Such differences are still issues in scientific research as well as points of
confusion for students.

43
 A material analogy relates structure in different material systems or
processes; for example, geometric similarity of a real car to a scale model of the
car. An important case that often goes unnoticed, because it is so subtle and
commonplace, is material equivalence of two material objects or systems,
whereby they are judged to be the same or identical. I call this an inductive
analogy, because it amounts to matching the objects to the same model (Fig. 4).
I submit that this matching process underlies classical inductive inference,
wherein repeated events are attributed to a single mechanism.
One other analogy deserves mention, because it plays an increasingly
central role in science: the analogy between conceptual models and computer
models. The formalization of mathematics has made it possible to imbed every
detail in the structure of conceptual models in computer programs, which,
running in simulation mode, can emulate the behavior of material systems with
stunning accuracy. More and more, computers carry out the empirical function
of matching models to data without human intervention. However there is an
essential difference between computer models and conceptual models, which
we discuss in the next section.
Considering the multiple, essential roles of analogy just described, I
recommend formalizing the concept of analogy in science with the technical
term MORPHISM. In mathematics a morphism is a structure-preserving
mapping: Thus the terms homomorphism (preserves algebraic structure) and
homeomorphism (preserves topological structure). Alternative notions of
analogy are discussed in [16].
The above characterization of science by Modeling Theory bears on deep
epistemological questions long debated by philosophers and scientists. For
example:
x In what sense can science claim objective knowledge about the material
world?
x To what degree is observed structure inherent in the material world and

(World 2) (World 3)
MENTAL WORLD CONCEPTUAL WORLD

Mental Models Understanding

Conceptual Models
)
(Subjective Creating (Objective)
Personal knowledge Scientific knowledge
Pe n encoded in
embodied rce tatio n symbolic
in neural
networks Ac pti
r pre ta tio forms
tio on e n
n Int pre
se
R e
Real Things
& Processes
Being and Becoming

PHYSICAL WORLD
(World 1)
Fig. 5. Mental models vs. Conceptual models 44
 independent of the observer?
x What determines the structure categories for conceptual models in Box
3?

In regard to the last question, I submit in line with Lakoff and Johnson [18, 19,
21] that these are basic categories of cognition grounded in the human sensory-
motor system. This suggests that answers to all epistemological questions
depend on our theory of cognition, to which we now turn.

VII. Modeling Structure of Cognition

If cognition in science is an extension of common sense, then the
structure of models in science should reflect structure of cognition in general.
To follow up this hint I outline a Modeling Theory of Cognition. The theory
begins with a crucial distinction between mental models and conceptual models
(Fig. 5). Mental models are private constructions in the mind of an individual.
They can be elevated to conceptual models by encoding model structure in
symbols that activate the individual’s mental model and corresponding mental
models in other minds. Just as Modeling Theory characterizes science as
construction and use of shared conceptual models, I propose to characterize
cognition as construction and manipulation of private mental models.
As already mentioned, the idea that mental models are central to
cognition is commonplace in cognitive science. However, it has yet to
crystallize into commonly accepted theory, so I cannot claim that other
researchers will approve of the way I construe their results as support for
Modeling Theory. The most extensive and coherent body of evidence comes
from cognitive linguistics, supporting the revolutionary thesis: Language does
not refer directly to the world, but rather to mental models and components
thereof! Words serve to activate, elaborate or modify mental models, as in
comprehension of a narrative.
This thesis rejects all previous versions of semantics, which located the
referents of language outside the mind, in favor of cognitive semantics, which
locates referents inside the mind. I see the evidence supporting cognitive
semantics as overwhelming [17-24], but it must be admitted that some linguists
are not convinced, and many research questions remain.
My aim here is to assimilate insights of cognitive linguistics into
Modeling Theory and study implications for cognition in science and
mathematics. The first step is to sharpen our definition of concept. Inspired by
the notion of ‘construction’ in cognitive
linguistics [25], I define a concept as a symbol
{form, meaning} pair represented by a
symbol (or symbolic construction), as CONCEPT
form meaning
schematized in Fig. 6. The meaning is
given by a mental model or schema Fig. 6: Concept triad
called a prototype, and the form is the
structure or a substructure of the prototype.

45
This is similar to the classical notion that the meaning of a symbol is given by
its intension and extension, but the differences are profound.
For example, the prototype for the concept right triangle is a mental
image of a triangle, and its form is a system of relations among its constituent
vertices and sides. The concept of hypotenuse has the same prototype, but its
form is a substructure of the triangle. This kind of substructure selection is
called profiling in cognitive linguistics. Note that different individuals can
agree on the meaning and use of a concept even though their mental images
may be different. We say that their mental images are homologous.
In my definition of a concept, the form is derived from the prototype.
Suppose the opposite. I call that a formal concept. That kind of concept is
common in science and mathematics. For example the concept of length is
determined by a system or rules and procedures for measurement that determine
the structure of the concept. To understand the concept, each person must
embed the structure in a mental model of his own making. Evidently formal
concepts can be derived from “informal concepts” by explicating the implicit
structure in a prototype. I submit that this process of explication plays an
important role in both developing and learning mathematics.
Like a percept, a concept is an irreducible whole, with gestalt structure
embedded in its prototype. Whereas a percept is activated by sensory input, a
concept is activated by symbolic input. Concepts can be combined to make
more elaborate concepts, for which I recommend the new term construct to
indicate that it is composed of irreducible concepts, though its wholeness is
typically than the “sum” of its parts.
We can apply the definition of ‘concept’ to sharpen the notion of
‘conceptual model,’ which was employed informally in the preceding section. A
conceptual model is now defined as a concept (or construct if you will) with the
additional stipulation that the structure of its referent be encoded in its
representation by a symbolic construction, or figure, or some other inscription.
Like a concept, a conceptual model is characterized by a triad, as depicted in
Fig. 7.
To emphasize the main point: the symbols for concepts refer to mental
models (or features thereof), which may or may not correspond to actual
material objects (as suggested in Fig.7). Though every conceptual model refers
to a mental model, the converse is not true. The brain creates all sorts of mental
constructions, including mental models, for which there are no words to
express. I refer to such constructions as ideas or intuitions. Ideas and intuitions
are elevated to concepts by creating symbols to represent them!

representation Material system

MODEL
structure referent Mental model

Fig. 7: Conceptual model

46
 My definitions of ‘concept’ and ‘conceptual model’ have not seen print
before, so others may be able to improve them. But I believe they incorporate
the essential ideas. The main task remaining is to elaborate the concept of
mental model with reference to empirical support for important claims.
The very idea of mental model comes from introspection, so that is a
good place to start. However, introspection is a notoriously unreliable guide
even to our own thinking, partly because most thinking is unconscious
processing by the brain. Consequently, like the tip of an iceberg, only part of a
mental model is open to direct inspection. Research has developed means to
probe more deeply.
Everyone has
imagination, the ability to Box 4: Spatial MENTAL models
conjure up an image of a x are schematic, representing only some
situation from a description features,
or memory. What can that x are structured, consisting of elements and
relations.
tell us about mental models?
x Elements are typically objects (or reified
Some people report images things).
that are picture-like, similar x Object properties are idealized (points,
to actual visual images. lines or paths).
However, others deny such x Object models are always placed in a
experience, and blind people background
are perfectly capable of x (context or frame).
imagination. Classical x Individual objects are modeled separately
from the frame, so they can move around in
research in this domain the frame.
found support for the view
that mental imagery is
internalized perception, but not without critics.
Barbara Tversky and collaborators [26] have tested the classical view by
comparison to mental model alternatives. Among other things, they compared
individual accounts of a visual scene generated from narrative with accounts
generated from direct observation and found that they are functionally
equivalent. A crucial difference is that perceptions have a fixed point of view,
while mental models allow change in point of view. Furthermore, spatial mental
models are more schematic and categorical than images, capturing some
features of the object but not all and incorporating information about the world
that is not purely perceptual. Major characteristics of spatial mental models are
summarized in Box 4. The best fit to data is a spatial framework model, where
each object has an egocentric frame consisting of mental extensions of three
body axes.
The general conclusion is that mental models represent states of the
world as conceived, not perceived. To know a thing is to form a mental model
of it. The details in Box 4 are abundantly supported by other lines of research,
especially in cognitive linguistics, to which we now turn.
In the preceding section we saw that concepts of structure and morphism
provide the foundation for models and modeling practices in science (and, later

47
I will claim, for mathematics as well). My purpose here is to link those concepts
to the extensive cognitive theory and evidence reviewed by Lakoff and
company [17-24], especially to serve as a guide for those who wish to mine the
rich lode of insight in this domain. To that end, I have altered Lakoff’s
terminology somewhat but I hope not misrepresented his message.
I claim that all reasoning is inference from structure, so I seek to
identify basic cognitive structures and understand how they generate the rich
conceptual structures of science and mathematics. The following major themes
are involved:
x Basic concepts are irreducible cognitive primitives grounded in sensory-
motor experience.
x All other conceptual domains are structured by metaphorical extension
from the basic domain.
x Cognition is organized by semantic frames, which provide background
structure for distinct conceptual domains and modeling in mental spaces.
Only a brief orientation to each theme can be given here.
Metaphors are morphisms in which structure in the source domain is
projected into the target domain to provide it with structure. The process begins
with grounding metaphors, which project structural primitives from basic
concepts. A huge catalog of metaphors has been compiled and analyzed to make
a strong case that all higher order cognition is structured in this way.
Semantic frames provide an overall conceptual structure linking systems
of related concepts (including the words that express them). In mathematics, the
frames may be general conceptual systems such as arithmetic and geometry or
subsystems thereof. Everyday cognition is structured by a great variety of
frames, such as the classic restaurant frame that that provides a context for
modeling what happens in a restaurant. A semantic frame for a temporal
sequence of events, such as dining (ordering, eating and paying for a meal), is
called a script.
Fauconnier has coined the term mental spaces for the arenas in which
mental modeling occurs [23, 24]. Especially significant is the concept of
blending, whereby distinct frames are blended to create a new frame. The
description of cognitive processes in such terms is in its infancy but very
promising.
As cognitive grounding for science and mathematics, we are most
interested in basic concepts of space, time and causality. Their prototypes,
usually called schemas, provide the primitive structures from which all
reasoning is generated. There are two kinds, called image schemas and
aspectual schemas.
Image schemas provide common structure for spatial concepts and
spatial perceptions, thus linking language with spatial perception. The world’s
languages use a relatively small number of image schemas, but they incorporate
spatial concepts in quite different ways –– in English mostly with prepositions.
Some prepositions, such as in/out and from/to, express topological concepts,
while others, such as up/down and left/right, express directional concepts.
The schema for each concept is a structured whole or gestalt, where in the

48
parts have no significance except in relation to the whole. For example, the
container schema (Fig. 8) consists of a boundary that separates interior and
exterior spaces. The preposition in profiles the interior, while out profiles the
exterior.
The container
schema provides the Container schema:
A
structure for the general Boundary A
concepts of containment in profiles Interior • in
and space as a container. out profiles Exterior • out
The alternative notion of Excluded middle: x in A or not in A
space as a set of points
(locations) was not Modus Ponens:
invented until the B A x in B ¼ x in A
nineteenth century. The • x Modus Tollens:
contrast between these two x not in A ¼ x not in B
concepts of space has
generated tension in the Fig. 8: Container Schema Logic
foundations of mathematics that is still not resolved to everyone’s satisfaction.
By metaphorical projection, the container schema structures many
conceptual domains. In particular, as Lakoff explains at length, the Categories-
are-containers metaphor provides
propositional logic with cognitive
grounding in the inherent logic of the source x
trajector goal
container schema (illustrated in Fig. 8). Fig. 9: Source-Path-Goal
More generally, container logic is the Schema
logic of part-whole structure, which
underlies the concepts of set and system (Box 4).
Aspectual schemas structure events and actions. The prototypical
aspectual concept is the verb, of which the reader knows many examples. The
most fundamental aspectual schema is the basic schema for motion (Fig. 9),
called the Source-Path-Goal schema by linguists, who use trajector as the
default term for any object moving along a path. This schema has its own logic,
and provides cognitive structure for the concepts of continuity and linear order
in mathematics. Indeed, Newton conceived of curves as traced out by moving
points, and his First Law of Motion provides grounding for the concept of time
on the more basic concept of motion [7]. Indeed, the Greek concept of a curve
as a locus of points suggests the action of drawing the curve. In physics the
concept of motion is integrated with concept of space, and the geometry of
motion is called kinematics.
Though the path schema of Fig. 9 is classified as aspectual in cognitive
linguistics, evidence from cognitive neuroscience and perceptual psychology
suggests that it should regarded as an image schema. It is a mistake to think that
visual processing is limited to static images. In visual cortex motion is
processed concurrently with form. Even young children can trace the path of a
thrown ball, and the path is retained mentally as a kind of afterimage, though,
like most of visual processing, it remains below the radar of consciousness.

49
 Clearly, the basic concepts of structure and quantity come from geometry.
Evidently the general concept of structure is derived from geometry by
metaphorical projection to practically every conceptual domain. An obvious
example is the general concept of state space, where states are identified with
locations.
Categories are fundamental to human thought, as they enable distinctions
between objects and events. One of the pillars of cognitive linguistics is Eleanor
Rosch’s discovery that Natural Categories are determined by mental
prototypes. This should be contrasted with the classical concept of a Formal
Category for which membership is determined by a set of defining properties, a
noteworthy generalization of the container metaphor. The notion of categories
as containers cannot account for a mountain of empirical evidence on natural
language use.
Natural categories (commonly called Radial categories) are discussed at
great length by Lakoff [18], so there is no need for details here. The term
“radial” expresses the fact that natural categories have a radial structure of
subordinate and superordinate categories with a central category for which
membership is determined by matching to a prototype. The matching process
accounts for fuzziness in category boundaries and graded category structure
with membership determined by partial matching qualified by hedges, such as
“It looks like a bird, but . . .”
The upshot is that the structure of natural categories is derived from
prototypes whereas for formal categories structure is imposed by conventions.
As already noted for formal concepts, formal categories play an essential role in
creating objective knowledge in science and mathematics. However, the role of
radial categories in structuring scientific knowledge has received little notice
[27].
Most human reasoning is inference from mental models. We can
distinguish several types of model-based reasoning:
x Abductive, to complete or extend a model, often guided by a semantic
frame in which the model is embedded.
x Deductive, to extract substructure from a model.
x Inductive, to match models to experience.
x Analogical, to interpret or compare models.
x Metaphorical, to infuse structure into a model.
x Synthesis, to construct a model, perhaps by analogy or blending other
models.
x Analysis, to profile or elaborate implicit structure in a model.
Justification of model-based reasoning requires translation from mental models
to inference from conceptual models that can be publicly shared, like the
scientific models in the preceding section.
In contrast, formal reasoning is computational, using axioms, production
rules and other procedures. It is the foundation for rigorous proof in
mathematics and formal logic. However, I daresay that mathematicians and
even logicians reason mostly from mental models. Model-based reasoning is
more general and powerful than propositional logic, as it integrates multiple

50
representations of information (propositions, maps, diagrams, equations) into a
coherently structured mental model. Rules and procedures are central to the
formal concept of inference, but they can be understood as prescriptions for
operations on mental models as well as on symbols.
We have seen how Modeling Theory provides a theoretical framework
for cognitive science that embraces the findings of cognitive linguistics. Thus it
provides the means for scientific answers to long-standing philosophical
questions, such as: What is the role of language in cognition? Is it merely an
expression of thought and a vehicle for communication? Or does it determine
the structure of thought? As for most deep philosophical questions, the answer
is “Yes and no!” Yes, the basic structure in thought is grounded in the evolved
structure of the sensory-motor system. No, there is more to the story. The
structure of mental models, perhaps even of aspectual and image schemas, is
shaped by experience with tools, linguistic as well as physical. In the following
sections we consider evidence for this in physics and mathematics.

VIII. Concepts of Force in science and common sense

From the beginning, Modeling Theory was developed with an eye to

improving instruction in science and mathematics, so we look to that domain for
validation of the theory. In section II, I reported the stunning success of
Malcolm Wells’ initial experiment with Modeling Theory and its subsequent
flowering in the Modeling Instruction Project. My purpose in this section is,
first to describe what Modeling Theory initially contributed to that success, and
second to propose new explanations based on the current version of the theory.
This opens up many opportunities for further research.
School physics has a reputation for being impossibly difficult. The rap is
that few have the talent to understand it. However, PER has arrived at a
different explanation by investigating common sense (CS) concepts of force and
motion in comparison to the Newtonian concepts of physics. The following
conclusions are now widely accepted:
x CS concepts dominate student thinking in introductory physics!
x Conventional instruction is almost totally ineffective in altering them!
x This result is independent of the instructor’s academic qualifications,
teaching experience, and (unless informed by PER) mode of teaching!
Definitive quantitative support for these claims was made possible by
development of the Force Concept Inventory (FCI). The initial results [3, 5]
have been repeatedly replicated (throughout the U.S. and elsewhere), so the
conclusions are universal, and only the ill-informed are skeptical.
The implications for conventional instruction could hardly be more
serious! Student thinking is far from Newtonian when they begin physics, and it
has hardly changed (<15%) when they finish the first course. Consequently,
students systematically misinterpret almost everything they read, hear and see
throughout the course. Evidence for this catastrophe has always been there for
teachers to see, but they lacked the conceptual framework to recognize it.

51
Witness the common student complaint: “I understand the theory, I just can’t
work the problems!” In my early years of teaching I dismissed such claims as
unfounded, because ability to work problems was regarded as the definitive test
of understanding. Now I see that the student was right. He did understand the
theory –– but it was the wrong theory! His theory wrapped up his CS concepts
in Newtonian words; he had learned jargon instead of Newtonian concepts.
Since students are oblivious to the underlying conceptual mismatch, they
cannot process their own mistakes in problem solving. Consequently, they
resort to rote learning and depend on the teacher for answers. A sure sign of this
state of affairs in a physics classroom is student clamoring for the teacher to
demonstrate solving more and more problems. They confuse memorizing
problem solutions with learning how to solve problems. This works to a degree,
but repeated failure leads to frustration and humiliation, self-doubt and
ultimately student turn-off!
Happily, this is not the end of the story. Figure 10 summarizes data from
a nationwide sample of 7500 high school physics students involved in the
Modeling Instruction Project during 1995–98. The mean FCI pretest score is
about 26%, slightly above the random guessing level of 20%, and well below
the 60% score which, for empirical reasons, can be regarded as a threshold in
the understanding of Newtonian mechanics.
Figure 10 shows that traditional high school instruction (lecture,
demonstration, and standard laboratory activities) has little impact on student
beliefs, with an average FCI posttest score of 42%, still well below the
Newtonian threshold. This is data from the classes of teachers before
participating in the Modeling Instruction Project.
Participating teachers attend an intensive 3-week Modeling Workshop
that immerses them in modeling pedagogy and acquaints them with curriculum
materials designed expressly to support it. Almost every teacher enthusiastically
adopts the approach and begins teaching with it immediately. After their first
year of teaching posttest scores for students of these novice modelers are about
10% higher, as shown in Fig. 10 for 3394 students of 66 teachers. Students of
expert modelers do much better.
For 11 teachers identified as expert modelers after two years in the
Project, posttest scores of their 647 students averaged 69%. Their average gain
is more than two standard deviations higher than the gain under traditional
instruction. It is comparable to the gain achieved by the first expert modeler
Malcolm Wells.
The 29%/69% pretest/posttest means for the expert modelers should
be compared with the 52%/63% means for calculus-based physics at a major
university [5]. We now have many examples of modelers who consistently
achieve posttest means from 80-90%. On the other hand, even initially under-
prepared teachers eventually achieve substantial gains, comparable to gains for
well-prepared teachers after two years in the project.
FCI scores are vastly more informative than scores for an ordinary test.
To see why, one needs to examine the structure of the test and the significance
of the questions. The questions are based on a detailed taxonomy of common

52
sense (CS) concepts
FCI mean
of force and motion score (%)
derived from
research. The
taxonomy is 80
Post-test
structured by a 69
60
systematic analysis of
the Newtonian force 52
40 42
concept into six
fundamental 26 26 29
Pre-test
20
conceptual
dimensions. Each Instruction
question requires a Traditional Novice Expert type
Modelers Modelers
forced choice FCI mean scores under different instruction types
between a Newtonian
concept and CS alternatives for best explanation in a common physical
situation, and the set of questions systematically probes all dimensions of the
force concept. Questions are designed to be meaningful to readers without
formal training in physics.
To a physicist the correct choice for each question is so obvious that the
whole test looks trivial. On the other hand, virtually all CS concepts about force
and motion are incompatible with Newtonian theory. Consequently, every
missed question has high information content. Each miss is a sure indicator of
non-Newtonian thinking, as any skeptical teacher can verify by interviewing the
student who missed it.
Considering the FCI’s comprehensive coverage of crucial concepts, the
abysmal FCI scores for traditional instruction imply catastrophic failure to
penetrate student thinking! Most high school students and half the university
students do not even reach the Newtonian threshold of 60%. Below that
threshold students have not learned enough about Newtonian concepts to use
them reliably in reasoning. No wonder they do so poorly on problem solving.
Why is traditional instruction so ineffective? Research has made the
answer clear. To cope with ordinary experience each of us has developed a
loosely organized system of intuitions about how the world works. That
provides intuitive grounding for CS beliefs about force and motion, which are
embedded in natural language and studied in linguistics and PER. Research
shows that CS beliefs are universal in the sense that they are much the same for
everyone, though there is some variation among individuals and cultures. They
are also very robust and expressed with confidence as obvious truths about
experience.
Paradoxically, physicists regard most CS beliefs about force and motion
as obviously false. From the viewpoint of Newtonian theory they are simply
misconceptions about the way the world truly is! However, it is more accurate,
as well as more respectful, to regard them as alternative hypotheses. Indeed, in
preNewtonian times the primary CS “misconceptions” were clearly articulated
and forcefully defended by great intellectuals –– Aristotle, Jean Buridan,

53
Galileo, and even Newton himself (before writing the Principia) [4]. Here we
see another side of the paradox:
To most physicists today Newtonian physics describes obvious structure in
perceptible experience, in stark contrast to the subtle quantum view of the
world. I have yet to meet a single physicist who recollects ever holding pre-
scientific CS beliefs, though occasionally one recalls a sudden aha! insight into
Newton’s Laws. This collective retrograde amnesia testifies to an important
fact about memory and cognition: recollections are reconstructed to fit current
cognitive structures. Thus, physicists cannot recall earlier CS thinking because
it is filtered by current Newtonian concepts.
In conclusion, the crux of the problem with traditional instruction is that it does
not even recognize CS beliefs as legitimate, let alone address them with
argument and evidence. In contrast, Modeling Instruction is deliberately
designed to address this problem with
x Modeling activities that systematically engage students in developing
models and providing their own explanations for basic physical
phenomena,
x Modeling discourse (centered on visual representations of the models) to
engage students in articulating their explanations and comparing them
with Newtonian concepts,
x Modeling concepts and tools (such as graphs, diagrams and equations) to
help students simplify and clarify their models and explanations.
Instructors are equipped with a taxonomy of CS concepts to help recognize
opportunities to elicit the concepts from students for comparison with
Newtonian alternatives and confrontation with empirical evidence. Instructors
know that students must recognize and resolve discrepancies by themselves.
Telling them answers does not work.
From years of
experimenting with Box 5 Contrasting Force Concepts
modeling discourse Posttest
(especially in the Newtonian vs. Common Sense Discrepancy
classroom of Malcolm
Wells) we have learned • First Law œ “Motion requires force” ~ 60%
to focus on the three (Impetus Principle)
CS concepts listed in
Box 5. When these • Second Law œ “Force is action” ~ 40%
(No Passive forces)
concepts are adequately
addressed, other • Third Law œ “Force is war” ~ 90%
misconceptions in our (Dominance Principle)
extensive taxonomy [5]
tend to fall away automatically. Their robustness is indicated by the posttest
discrepancies (Box 5) from FCI data on more than a thousand university
students. After completing a first course in calculus-based physics, the fraction
of students choosing CS alternatives over Newton’s First, Second and Third
Laws was 60%, 40% and 90% respectively. Of course, Newton’s Laws are not

54
named as such in the FCI. 80% of the students had already taken high school
physics and could state Newton’s Laws as slogans before beginning university
physics.
After the Modeling Instruction Project was up and running, I learned
about Lakoff’s work on metaphors and its relevance for understanding CS force
and motion concepts. I presented the ideas to teachers in Modeling Workshops
but have no evidence that this improved the pedagogy, which was already well
developed. I suppose that much of the new insight was overlooked, because it
was not nailed down in print, so let me record some of it here as analysis of the
three primary CS concepts in Box 5.
The Impetus Principle employs the Object-As-Container metaphor, where
the container is filled with impetus that makes it move. After a while the
impetus is used up and the motion stops. Of course, students don’t know the
term impetus (which was coined in the middle ages); they often use the term
energy instead. Naïve students don’t discriminate between energy and force.
Like Newton himself before the Principia, they have to be convinced that “free
particle motion in a straight line” is a natural state that doesn’t require a motive
force (or energy) to sustain it. This does not require discarding the impetus
intuition (which is permanently grounded in the sensory-motor system in any
case) but realigning the intuition with physics concepts of inertia and
momentum.
The CS prototype for force is human action on an object.
Consequently, students don’t recognize constraints on motion like walls and
floors as due to contact forces. “They just get in the way.” Teachers try to
activate student intuition by emphasizing that “force is a push or a pull,”
without realizing that unqualified application of this metaphor excludes passive
forces. Besides, no textbooks explicitly note that universality of force is an
implicit assumption in Newtonian theory, which requires that motion is
influenced only by forces. To arrive at force universality on their own, students
need to develop intuition to recognize forces in any instance of physical contact.
As an instructional strategy to achieve that end, Clement and Camp [28] engage
students in constructing a series of “bridging analogies” to link, for example, the
unproblematic case of a person pressing on a spring to the problematic case of a
book resting on a table. I recommend modifying their approach to include a
common vector representation of normal force in each case to codify symbolic
equivalence (as in Fig. 4).
In situations involving Newton’s Third Law, the slogan “for every action
there is an equal and opposite reaction” evokes a misplaced analogy with a
struggle between “opposing forces,” from which it follows that one must be the
winner, “overcoming” the other, in contradiction to the Third Law. The
difficulty that students have in resolving this paradox is reflected in the fact that
FCI questions on the Third Law are typically the last to be mastered. DiSessa
[29] gives a perceptive analysis of Third Law difficulties and measures to
address them.
Such insights into student thinking as just described are insufficient for
promoting a transition to Newtonian thinking in the classroom. The literature is

55
replete with attempts to address specific misconceptions with partial success at
best. So what accounts for the singular success of Modeling Instruction as
measured by the FCI (Fig. 10)? As for any expert performance, detailed
planning and preparation is essential for superior classroom instruction. (The
intensive Modeling Workshops help teachers with that.) However, Modeling
Instruction is unique in its strategic design.
Rather than address student misconceptions directly, Modeling
Instruction creates an environment of activities and discourse to stimulate
reflective thinking about physical phenomena that are likely to evoke those
misconceptions. The environment is structured by an emphasis on models and
modeling with multiple representations (maps, graphs, diagrams, equations).
This provides students with conceptual tools to sharpen their thinking and gives
them access to Newtonian concepts. In this environment students are able to
adjust their thinking to resolve discrepancies within the Newtonian system,
which gradually becomes their own. Rather than learning Newtonian concepts
piecemeal, they learn them as part of a coherent Newtonian system.
Construction of a Newtonian model requires coordinated use of all the
Newtonian concepts, and only this reveals the coherence of the Newtonian
system. That coherence is not at all obvious from the standard statement of
Newton’s Laws. I believe that learning Newtonian concepts as a coherent
system best accounts for high FCI scores. Logically this is only a sufficient
condition for a high score, but I estimate that a high score from piecemeal
understanding of Newtonian physics is improbably low. Thus, it is best to
interpret overall FCI score as a measure of coherence in understanding
Newtonian physics.
One other important point deserves mention here. As we have noted,
Modeling Theory informed by empirical evidence from cognitive science holds
that mental models are always constructed within a semantic frame.
Accordingly, I suppose that physical situations (regardless of how they are
presented) activate a Newtonian semantic frame in the mental spaces of
physicists. And I submit that physics instruction is not truly successful until the
same is true for students. It is well known that students tend to leave the science
they have learned in the classroom and revert to CS thinking in every day
affairs. Perhaps recognizing this as a problem of semantic framing can lead to a
better result.
As I have described it, Modeling Instruction does not depend on detailed
understanding of how students think. Indeed, I have tried to steer it clear of
doubtful assumptions about cognition that might interfere with learning.
However, I now believe that advances in the Modeling Theory of cognition
described in Section VII are sufficient to serve as a reliable guide for research to
further improve instruction by incorporating details about cognition. Let me
sketch the prospects with specific reference to force and motion concepts.
The intertwined concepts of force and causation have been studied
extensively in cognitive linguistics. Lakoff and Johnson [19] show that the great
variety of causal concepts fall naturally into a radial category (“kinds of
causation”) structured by a system of metaphorical projections. The central

56
prototype in this category is given by the Force-as-Human-Action metaphor, in
agreement with our analysis above. Their analysis provides an organizational
framework for the whole body of linguistic research on causation. That research
provides valuable insight into CS concepts of force and motion that deserves
careful study. However, limited as it is to study of natural languages, linguistic
research does not discover the profound difference in the force concept of
physicists. For that we need to turn to PER, where the deepest and most
thorough research is by Andy diSessa [29].
In much the same way that linguists have amassed evidence for the
existence of prototypes and image schemas, diSessa has used interview
techniques to isolate and characterize conceptual primitives employed by
students in causal reasoning. He has identified a family of irreducible
“knowledge structures” that he calls phenomenological primitives or p-prims.
Since diSessa’s definitive monograph on p-prims in 1993, converging evidence
from cognitive linguistics has made it increasingly clear that his p-prims are of
the same ilk as the image and aspectual schemas discussed in the preceding
section. Accordingly, I aim to integrate them under the umbrella of Modeling
Theory.
Let us begin with the most important example, which diSessa calls Ohm’s
p-prim. As he explains,
Ohm’s p-prim comprises “an agent that is the locus of an impetus
that acts against a resistance to produce a result.”
Evidently this intuitive structure is abstracted from experience pushing objects.
It is an important elaboration of the central Force-as-Action metaphor
mentioned above –– Very important! –– Because this structure is fundamental
to qualitative reasoning. The logic of Ohm’s p-prim is the qualitative
proportion:
more effort Ÿ more result,
and the inverse proportion:
more resistance Ÿ less result.
This reasoning structure is evoked for explanatory purposes in circumstances
determined by experience.
DiSessa identifies a number of other p-prims and catalogs them into a
cluster that corresponds closely to the taxonomy of CS force and motion
concepts used to construct the FCI. His monograph should be consulted for
many details and insights that need not be repeated here. Instead, I comment on
general aspects of his analysis.
In accord with Lakoff and Johnson, diSessa holds that causal cognition is
grounded in a loosely organized system of many simple schemas derived from
sensory-motor experience. P-prims provide the grounding for our intuitive sense
of (causal) mechanism. They are the CS equivalent of physical laws, used to
explain but not explainable. To naïve subjects, “that’s the way things are.”
As to be expected from their presumed origin in experience, p-prims are
cued directly by situations without reliance on language. DiSessa asserts that p-

57
prims are inarticulate, in the sense that they are not strongly coupled to
language. Here there is need for further research on subtle coupling with
language that diSessa has not noticed. For example, Lakoff notes that the
preposition on activates and profiles schemas for the concepts of contact and
support, which surely should be counted among the p-prims.
As disclosed in Ohm’s p-prim, the concept of (causal) agency entails a
basic
Causal syntax: agent o (kind of action) o on patient o result.
DiSessa notes that this provides an interpretative framework for F = ma, and he
recommends exploiting it in teaching mechanics. However he does not
recognize it as a basic aspectual schema for verb structure, which has been
studied at length in cognitive grammar [22]. Aspectual concepts are generally
about event structure, where events are changes of state and causes (or causal
agents) induce events. Causes cannot be separated from events. Here is more
opportunity for research.
Under physics instruction, diSessa says that p-prims are refined but not
replaced, that they are gradually tuned to expertise in physics. Considering the
role of metaphor and analogy in this process, it might be better to say that p-
prims are realigned. There are many other issues to investigate in this domain.
Broadly speaking, I believe that we now have sufficient theoretical resources to
guide research on instructional designs that target student p-prims more directly
to retune and integrate them into schemas for more expert-like concepts. I
propose that we design idealized expert prototypes for force and motion
concepts to serve as targets for instruction. This would involve a more targeted
role for diagrams to incorporate figural schemas into the prototypes.
The call to design expert prototypes embroils us in many deep questions
about physics and epistemology. For example, do forces really exist outside our
mental models? We have seen that Modeling Theory tells us that the answer
depends on our choice of theoretical primitives and measurement conventions.
Indeed, if momentum is a primitive, then Newton’s Second Law is reduced to a
definition of force as momentum flux and the Third Law expresses momentum
conservation. The physical intuition engaged when mechanics is reformulated in
terms of momentum and momentum flux has been investigated by diSessa
among others, but few physicists have noted that fundamental epistemological
issues are involved. Not the least of these issues is the transition from classical
to quantum mechanics, where momentum is king and force is reduced to a
figure of speech.
A related epistemological question: Is causal knowledge domain-
specific? Causal claims are supported by causal inference from models based on
acquired domain-specific knowledge. But to what degree does inference in
different domains engage common intuitive mechanisms? Perhaps the
difference across domains is due more to structure of the models rather than the
reasoning. Perhaps we should follow Lakoff’s lead to develop force and
interaction as a radial category for a progression of interaction concepts ranging
from particles to fields.

58
 I am often asked how the FCI might be emulated to assess student
understanding in domains outside of mechanics, such as electrodynamics,
thermodynamics, quantum mechanics and even mathematics. Indeed, many
have tried to do it themselves, but the result has invariably been something like
an ordinary subject matter test. The reason for failure is insufficient attention to
cognitive facts and theory that went into FCI design, which I now hope are
more fully elucidated by Modeling Theory. The primary mistake is to think that
the FCI is basically about detecting misconceptions in mechanics. Rather, as we
have seen, it is about comparing CS causal concepts to Newtonian concepts.
The p-prims and image schemas underlying the CS concepts are not peculiar to
mechanics, they are basic cognitive structures for reasoning in any domain.
Therefore, the primary problem is to investigate how these structures are
adapted to other domains. Then we can see whether reasoning in those domains
requires other p-prims that have been overlooked. Finally, we can investigate
whether and how new p-prims are created for advanced reasoning in science
and mathematics. That brings us to the next section, where we discuss the
development of conceptual tools to enhance scientific thinking.

IX. Tools to think with

The evolution of science is driven by invention and use of tools of increasing

sophistication and power! The tools are of two kinds: instruments for detecting
patterns in the material world, and symbolic systems to represent those patterns
for contemplation. As outlined in Fig. 11, we can distinguish three major stages
in tool development.
In the perceptual domain, pattern detection began with direct observation
using human sensory apparatus. Then the perceptual range was extended by
scientific instruments such as telescopes and microscopes. Finally, human
sensory detectors are replaced by more sensitive detection instruments, and the
data are processed by computers with no role for humans except to interpret the
final results; even there the results may be fed to a robot to take action with no
human participation at all.
Tool development in the cognitive domain began with the natural
languages in spoken and then written form. Considering their ad hoc evolution,
the coherence, flexibility and subtlety of the natural languages is truly
astounding. More deliberate and systematic development of symbolic tools
came with the emergence of science and mathematics. The next stage of
enhancing human cognitive powers with computer tools is just beginning. My
purpose in this section is to discuss what Modeling Theory can tell us about the
intuitive foundations of mathematics to serve as a guide for research on design
of better instruction and better mathematical tools for modeling in science and
engineering.

59
 Modeling Tool Development

Instruments Symbolic systems

for detecting patterns precision for representing patterns

Perceptual Stage I Cognitive

• Sensory apparatus • natural language
(direct observation) • sketches & icons
Experimental Stage II Theoretical
• scientific instruments • mathematics
& apparatus • diagrammatic systems
Computer Stage III Computer
• Universal Lab • simulation
Interface • visualization
Modeling skills
Fig. 11

While science is a search for structure, mathematics is the science of

structure. Every science develops specialized modeling tools to represent the
structure it investigates. Witness the rich system of diagrams that chemists have
developed to characterize atomic and molecular structure. Ultimately, though,
these diagrams provide grist for mathematical models of greater explanatory
power. What accounts for the ubiquitous applicability of mathematics to
science?
I have long wondered how mathematical thinking relates to theoretical
physics. According to Modeling Theory, theoretical physics is about designing
and analyzing conceptual models that represent structure in the material world.
For the most part these models are mathematical models, so the cognitive
activity is called mathematical modeling. But how does mathematical thinking
differ from the mathematical modeling in physics? Can it be essentially the
same when there are no physical referents for the mathematical structures? I am
now convinced that the answer is yes! The light went on when I learned about
cognitive semantics and realized that the referents for cognition in both
mathematics and physics are mental models! Lakoff and Núñez [20] argue
forcefully for the same conclusion, but I want to put my own twist on it.
I contend that the basic difference between mathematics and physics is
how they relate their mental models to the external world. Physicists aim to
match their mental models to structure in the material world. I call the ability to
make such matches physical intuition. Note that mathematics is not necessarily

60
involved in this. In contrast, mathematicians aim to match their mental models
to structure in symbolic systems. I call the ability to make such matches
mathematical intuition. To be sure, physicists also relate their mental models
to mathematical structures, but for the most part they take the mathematics as
given. When they do venture to modify or extend the mathematical structures
they function as mathematicians. Indeed, that is not uncommon; a vast portion
of mathematics was created by theoretical physicists.
According to Modeling Theory, mathematicians work with intuitive
structures (grounded in sensory-motor experience) that every normal person
has. They proceed to encode these structures in symbolic systems and elaborate
them using the intuitive inferential structures of p-prims and image schemas. I
submit that mathematical thinking involves a feedback loop generating external
symbolic structures that stimulate modeling in mental spaces to generate more
symbolic structure. Though some mathematical thinking can be done with
internal representations of the symbols, external representation is essential for
communication and consensus building [30]. For this reason, I believe that the
invention of written language was an essential prerequisite to the creation of
mathematics.
Let’s consider an example of intuitive grounding for mathematical structures.
Lakoff and Núñez [20] give many others, including four grounding metaphors
for arithmetic. Note that the intuitive causal syntax discussed in the previous
section can be construed (by metaphorical projection at least) as
Operator syntax: agent o (kind of action) o on patient o result,
where the action is on symbols (instead of material objects) to produce other
symbols. Surely this provides an intuitive base for the mathematical concept of
function (though it may not be the only one). Exploration of mental models
reveals various kinds of structure that can be encoded and organized into
symbolic systems such as Set theory, Geometry, Topology, Algebra and Group
theory. Note that the number of distinct types of mathematical structure is
limited, which presumably reflects constraints on their grounding in the
sensory-motor system. Of course, to confirm this point of view thorough
research is needed to detail the intuitive base for each type of mathematical
structure. Lakoff and Núñez [20] have already made a good start.
The upshot is that cognitive processes in theoretical physics and
mathematics are fundamentally the same, centered on construction and analysis
of conceptual models. Semantics plays a far more significant role in
mathematical thinking (and human reasoning in general) than commonly
recognized –– it is the cognitive semantics of mental models, mostly residing in
the cognitive unconscious, but often manifested in pattern recognition and
construction skills [31]. Mathematical intuition (like physical intuition) is a
repertoire of mental structures (schemas) for making and manipulating mental
models! This goes a long way toward answering the question: What does it
mean to understand a scientific concept?

61
I am not alone in my opinion on the intimate relation between physics and
mathematics. Here is a brief extract from a long diatribe On Teaching
Mathematics by the distinguished Russian mathematician V. I. Arnold [32]:
“Mathematics is a part of physics. Physics is an experimental science, a part
of natural science. Mathematics is the part of physics where experiments are
cheap. . . . In the middle of the 20th century it was attempted to divide
physics and mathematics. The consequences turned out to be catastrophic.
Whole generations of mathematicians grew up without knowing half of their
science and, of course in total ignorance of other sciences.”
Arnold is deliberately provocative but not flippant. He raises a very important
educational issue that deserves mention quite apart from the deep connection to
cognitive science that most concerns us here.
There is abundant evidence to support Arnold’s claim. For example, up
until World War II physics was a required minor for mathematics majors in US
universities. Since it was dropped, the mathematics curriculum has become
increasingly irrelevant to physics majors, and physics departments provide most
of the mathematics their students need. At the same time, mathematicians have
contributed less and less to physics, with some exceptions like the Russian
tradition that Arnold comes from, which has sustained a connection to physics.
But the most serious consequence of the divorce of mathematics from physics is
the fact that, in the U.S. at least, most high school math teachers have little
insight into relations of math they teach to science in general and physics in
particular. Here is a bit of data to support my contention: We administered the
FCI to a cohort of some 20 experienced high school math teachers. The profile
of scores was the same as the pitiful profile for traditional instruction in Fig. 10,
with the highest score at the Newtonian threshold of 60%. Half the teachers
missed basic questions about relating data on motion to concepts of velocity and
acceleration. This chasm between math and science, now fully ensconced in the
teachers, may be the single most serious barrier to significant secondary science
education reform.
To document deficiencies in math education, many have called for a
Math Concept Inventory (MCI) analogous to the FCI. I have resisted that call
for lack of adequate theory and data on intuitive foundations for mathematical
thinking. There is lots of educational research on conceptual learning in
mathematics, but most of it suffers from outdated cognitive theory. Modeling
Theory offers a new approach that can profit immediately from what has been
learned about cognitive mechanisms in physics. We need to identify “m-prims”
that are mathematical analogs of the p-prims discussed in the preceding section.
I suspect that underlying intuitive mechanisms are the same for m-prims and p-
prims, but their connections to experience must be different to account for the
difference between mathematical and physical intuition noted above. I
recommend coordinated research on m-prims and p-prims aiming for a
comprehensive Modeling Theory of cognition in science and mathematics.
I have barely set the stage for application of Modeling Theory for my
favorite enterprise, namely, the design of modeling tools for learning and doing
science, engineering and mathematics [10]. I have previously described the

62
influence of my Geometric Algebra research on development of Modeling
Theory [13]. Now I believe that Modeling Theory has matured to the point
where it can contribute, along with Geometric Algebra, to the design of more
powerful modeling tools, especially tools embedded in computer software. But
that is a task for tomorrow!
X. Conclusion

Central thesis: Cognition in science, mathematics, and everyday life is

basically about making and manipulating mental models.
• The human cognitive capacity for creating, manipulating and
remembering mental models has evolved to facilitate coping with the
environment, so it is central to “common sense” thinking and
communication by humans.
• Human culture has expanded and augmented this capacity by creating
semiotic systems: representational systems of signs (symbols, diagrams,
tokens, icons, etc.), most notably spoken and written language.
• Science and mathematics has further extended the use of symbolic
systems deliberately and self-consciously, but the cognitive mechanisms
involved are essentially the same as for common sense.

Scientific modeling is a “deliberate and self-conscious extension of the evolved

cognitive capabilities for “mapping” the environment.” (Giere)

Science is a refinement of common sense! differing in respect to:

Objectivity – based on explicit rules & conventions for observer-
independent inferences

Precision – in measurement
– in description and analysis
Formalization – for mathematical modeling and analysis of
complex systems
Systematicity – coherent, consistent & maximally integrated bodies
of knowledge
Reliability – critically tested & reproducible results
Skepticism – about unsubstantiated claims
Knowledge and Wonder – so say Weisskopf & Sagan
Social structure and norms – Ziman

References

[1] Geometric Calculus Research: <http://modelingnts.la.asu.edu>

63
[2] D. Hestenes (1979), Wherefore a Science of Teaching, The Physics Teacher
17: 235-242.
[3] I. Halloun and D. Hestenes (1985), Initial Knowledge State of College
Physics Students, Am. J. Phys. 53: 1043-1055.
[4] I. Halloun and D. Hestenes(1985), Common Sense Concepts about Motion,
Am. J. Phys. 53, 1056-1065.
[5] D. Hestenes, M. Wells, and G. Swackhamer (1992), Force Concept
Inventory, Physics Teacher 30: 141-158.
[6] D. Hestenes, Toward a Modeling Theory of Physics Instruction (1987), Am.
J. Phys. 55: 440-454.
[7] D. Hestenes, Modeling Games in the Newtonian World, Am. J. Phys. 60:
732-748 (1992).
[8] M. Wells, D. Hestenes, and G. Swackhamer (1995), A Modeling Method for
High School Physics Instruction, Am. J. Phys. 63: 606-619.
[9] D. Hestenes (1997), Modeling Methodology for Physics Teachers. In E.
Redish & J. Rigden (Eds.) The changing role of the physics department in
modern universities. American Institute of Physics Part II. p. 935-957.
[10] D. Hestenes (1996), Modeling Software for learning and doing physics. In
C. Bernardini, C. Tarsitani and M. Vincentini (Eds.), Thinking Physics for
Teaching, Plenum, New York, p. 25-66.
[11] Modeling Instruction Project: <http://modeling.asu.edu> This website
includes research articles as well as details about implementation and
evaluation.
[12] D. Hestenes (1999), New Foundations for Classical Mechanics, (Reidel,
Dordrecht/Boston, 1986), Second Edition (Kluwer, Dordrecht/Boston).
[13] D. Hestenes (2003), Oersted Medal Lecture 2002: Reforming the
Mathematical Language of Physics, Am. J. Phys. 71:104-121.
[14] D. Hestenes (2005), Gauge Theory Gravity with Geometric Calculus,
Foundations of Physics 33: 903-970.
[15] L. Magnani, N. Nercessian & P. Thagard (1999), Model-Based Reasoning
in Scientific Discovery (Kluwer Academic/Plenum Publishers, New York).
[16] D. Gentner, K. Holyoak & B. Kokinov (2001), The Analogical Mind (MIT
Press, Cambridge MA).
[17] G. Lakoff & M. Johnson (1980), Metaphors We Live By (U. Chicago Press,
Chicago)
[18] G. Lakoff (1987), Women, Fire and Dangerous Things: what categories
reveal about the mind (U. Chicago Press, Chicago)
[19] G. Lakoff & M. Johnson (1999), Philosopy in the Flesh (Basic Books, New
York)
[20] G. Lakoff & R. Núñez (2000), Where mathematics comes from. (Basic
Books, New York)
[21] M. Johnson (1987), The Body in the Mind: The Bodily Basis of Meaning,
Imagination, and reason. (U. Chicago Press, Chicago)
[22] R. Langacker (2000), Grammar and Conceptualization (Mouton de
Gruyter, Berlin).
[23] G. Fauconnier (1997), Mappings in Thought and Language. (Cambridge .

64
 Press, New York).
[24] G. Fauconnier & M. Turner (2002), The Way We Think: conceptual
blending and the mind’s hidden complexities. (Basic Books, New York).
[25] A. Goldberg (1995), Constructions: A construction grammar approach to
argument structure. (U. Chicago Press, Chicago)
[26] Bryant, Tversky & Lanca (2000), Retrieving Spatial Relations from
Observation and Memory. In van der Zee & Nikanne (eds.), Cognitive
Interfaces (Oxford U. Press, Oxford)
[27] R. Giere (1999). SCIENCE without LAWS (U. Chicago Press, Chicago).
[28] J. Clement & C. Camp (1994), Preconceptions in Mechanics
(Kendall/Hunt, Dubuque).
[29] A. diSessa (1993), Toward an Epistemology of Physics, Cognition &
Instruction 10: 105-225.
[30] J. Ziman (1978). Reliable knowledge, An exploration of the grounds for
belief in science (Cambridge U. Press, Cambridge).
[31] P. Ross (2006), The Expert Mind, Scientific American 295: 64-71.
[32] V. I. Arnold (1997), On Teaching Mathematics,
(http://pauli-uni-muenster.de/~munsteg/arnold.html)

65