Вы находитесь на странице: 1из 27

This article was downloaded by: [New York University]

On: 23 April 2013, At: 12:18


Publisher: Routledge
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House,
37-41 Mortimer Street, London W1T 3JH, UK

Educational Psychologist
Publication details, including instructions for authors and subscription information:
http://www.tandfonline.com/loi/hedp20

Knowledge structure and problem solving in physics


a

F. Reif & Joan I. Heller

Physics Department and Group in Science and Mathematics Education, University of California,
Berkeley, California, 94720
b

University of California, Berkeley


Version of record first published: 01 Oct 2009.

To cite this article: F. Reif & Joan I. Heller (1982): Knowledge structure and problem solving in physics, Educational Psychologist,
17:2, 102-127
To link to this article: http://dx.doi.org/10.1080/00461528209529248

PLEASE SCROLL DOWN FOR ARTICLE


Full terms and conditions of use: http://www.tandfonline.com/page/terms-and-conditions
This article may be used for research, teaching, and private study purposes. Any substantial or systematic
reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in any form to anyone
is expressly forbidden.
The publisher does not give any warranty express or implied or make any representation that the contents
will be complete or accurate or up to date. The accuracy of any instructions, formulae, and drug doses should
be independently verified with primary sources. The publisher shall not be liable for any loss, actions, claims,
proceedings, demand, or costs or damages whatsoever or howsoever caused arising directly or indirectly in
connection with or arising out of the use of this material.

Educational Psychologist
1982, Vol. 17, No. 2, 102127

Knowledge Structure
and Problem Solving in Physics
F. Reif and Joan I. Heller

Downloaded by [New York University] at 12:18 23 April 2013

University of California, Berkeley


This article presents a prescriptive analysis of the kinds of knowledge and
procedures leading to effective human problem solving in a quantitative
science such as physics. The knowledge about such a science, explicated in
the case of mechanics, specifies special descriptive concepts and relations
described at various levels of abstractness, is organized hierarchically, and is
accompanied by explicit guidelines specifying when and how this knowledge
is to be applied. General problem-solving procedures, to be used in conjunction with such domain-specific knowledge, specify how initially to describe
and analyze any problem so as to facilitate its subsequent solution; how to
search for a solution by methods of constraint satisfaction used together with
heuristic methods for decomposing problems and exploring decisions; and
how to assess whether the resulting solution is correct and reasonably optimal. The preceding model of effective human problem solving is compared
with some relevant observations and with special experiments designed to
test such a prescriptive model. It also suggests methods for teaching students
improved scientific problem-solving skills.
also revealed that the knowledge structures
and problem-solving procedures of experts and
novices differ in significant ways.
Although studies of the thought processes of
experts and novices have yielded valuable insights about problem solving, they have
significant limitations. For example, it is unwise to assume that experts necessarily perform
optimally. Furthermore, educational efforts
must do more than merely teach students to
perform as experts do. Instead, students must
often be taught to use explicit procedures to
accomplish tasks which experts perform almost
automatically because they recognize familiar
patterns as a result of years of experience.
The work described in this article has, therefore, aimed to study human problem solving
from a more general point of view which
transcends the investigation of naturally occuring intellectual functioning. In particular, we
have sought to answer the following general
question: What kinds of knowledge and procedures enable human subjects to achieve good
problem-solving performance in a realistic
scientific domain?
This question is "prescriptive" and more
general than a descriptive concern with
naturalistic phenomena. Thus it may be
answered without necessarily trying to simulate
what actual experts do and without assuming
that experts always perform optimally. Of
course, a prescriptive theoretical model of

Recent years have witnessed increasing interest


in understanding human cognitive processes in
realistically complex domains. In particular,
while earlier studies of human problem solving
dealt largely with puzzles and games (Newell &
Simon, 1972), more recent work has focused attention on problem solving in complex domains
of practical educational or scientific interest. For
example, such work has been pursued in
mathematics (Greeno, 1978a, 1978b), engineering thermodynamics (Bhaskar & Simon, 1977),
electronics (Brown & Burton, 1975), and physics
(Chi, Feltovich, & Glaser, 1981; Chi, Glaser, &
Rees, 1981; Larkin, McDermott, Simon, &
Simon, 1980; Larkin & Reif, 1979; Simon &
Simon, 1978). Furthermore, work in artificial intelligence has also been extended to scientific domains such as physics (Bundy, 1978; DeKleer,
1977; Novak, 1977), chemistry (Feigenbaum,
1977), and others.
Such work has shown that effective problem
solving in a realistic domain depends crucially
on the content and structure of the knowledge
about the particular domain. The studies have
This work was supported in part by the National Science
Foundation under grant #SED79-20592. We wish to thank
Allan Collins for several useful discussions and for playing
so well the role of an obliging naive "robot" solving unfamiliar physics problems under external control.
The address of F. Reif is Physics Department and Group
in Science and Mathematics Education; University of
California; Berkeley, California 94720.

Copyright 1982 by Division 15 of the American Psychological Association, Inc.

102

Downloaded by [New York University] at 12:18 23 April 2013

KNOWLEDGE STRUCTURE AND PROBLEM SOLVING

good performance may be suggested by


naturalistic observations of expert behavior;
however, it may also be suggested by purely
theoretical task analyses. Correspondingly, the
sole criterion of validity of such a prescriptive
model is that it lead to predictably effective
performance when implemented by a human
subject, even if it does not mimic what experts do.
. To clarify the distinction between prescriptive and descriptive points of view, imagine
that a hypothetical cognitive scientist, working
in the year 100 AD, had been trying to formulate a model of good performance in
arithmetical problem solving. If the model had
suggested the use of the modern place-value
representation of numbers, it would have led
to very good arithmetical performance and
would thus have been an excellent prescriptive
model. However, it would have been an unsatisfactory descriptive model of the behavior
of experts, all of whom used Roman numerals
at that time.
A prescriptive perspective is prevalent in artificial intelligence where the aim is to create
programs that can perform intellectual tasks
effectively, without any necessary relation to
human cognition. But, as pointed out
elsewhere (Reif, 1979), a prescriptive approach
can also be of major interest in work on human
cognitive processes. Thus, it can be very useful
for identifying essential knowledge required
for good performance and can thereby help to
make explicit expert knowledge which is often
largely tacit. Furthermore, such an approach is
of essential importance in any applied work
aiming to improve human performance, to
design educationally effective instruction, or to
exploit the potentialities of systems involving
person-computer interaction.
In trying to specify the knowledge and procedures leading to good human performance
in a scientific domain, we have focused our attention on problem-solving performance in
basic college-level physics, specifically in the
field of mechanics. This scientific domain is
realistically complex, requires flexible problem
solving based on a considerable amount of
special knowledge, is often difficult for many
students, and is representative of other quantitative sciences or engineering fields. On
the other hand, this domain is sufficiently simple to permit an analysis of underlying
cognitive processes. Furthermore, such an
analysis can draw upon insights derived from

103

previous observations of experts and novices in


this domain (e.g., Larkin & Reif, 1979).
Our work has presupposed that the human
subjects, who engage in scientific problemsolving tasks, possess well-developed human
capabilities, such as natural-language
understanding, basic mathematical and
algebraic skills, and the ability to draw
and interpret diagrams. On the other hand, we
do not assume that these subjects possess more
sophisticated or strategic forms of knowledge
required for scientific problem solving. Indeed, our main aim has been precisely to
specify explicitly these more sophisticated
forms of knowledge leading to good problemsolving performance.
The problem-solving model discussed in the
following sections encompasses some general
problem-solving procedures to be used in conjunction with a knowledge base about a particular scientific domain. The procedures
include procedures for initially describing and
analyzing any problem so as to bring it into a
form facilitating its subsequent solution, procedures for constructing the solution by
methods and decision processes facilitating
search, and finally procedures for assessing
whether the resulting solution is correct and
reasonably optimal. The content and organization of the domain-specific knowledge base are
specifically designed to facilitate the implementation of these general problem-solving
procedures.
In the following pages we shall first discuss
the knowledge base, exhibited in the specific
domain of mechanics, since this will allow us
to illustrate subsequent general remarks with
concrete examples. Then we shall turn our attention to the general problem-solving procedures and illustrate how they are used in
conjunction with the knowledge base. Finally,
we shall mention briefly some experimental
tests of this model and point out some of its
practical educational implications.
We shall strive to provide sufficient details
and examples to make clear the nature of the
knowledge and cognitive processes contributing to effective human problem solving.
On the other hand, we shall try to avoid excessive details that might be distracting or of
predominant interest to practicing physicists.
Thus the discussion is intended to be
understandable to non-technical readers even
if they cannot fully interpret specific
equations.

Downloaded by [New York University] at 12:18 23 April 2013

104

F. REIF AND JOAN I. HELLER

KNOWLEDGE BASE

Top Level of the Knowledge Base

Problem solving in a complex scientific domain is impossible without a knowledge base


containing substantial amounts of specialized
knowledge about this domain. It is important
to specify not only the content of this
knowledge base, but also how this knowledge
is organized and symbolically represented to
facilitate efficient information retrieval and
problem solving. Accordingly, we shall devote
this section of the article to propose a particular structure of the knowledge base for the
chosen prototype domain of mechanics. This
structure is not found in any conventional textbook, but is specifically designed to aid effective problem solving and resembles somewhat
the knowledge possessed by experts.
The knowledge base discussed in the following pages is organized so as to facilitate information retrieval by progressively narrowing the
domain of search, starting from gross information and proceeding to successively more
detailed information with relatively few decisions at each step. Thus the entire knowledge
is subdivided into knowledge blocks, of a size
conveniently small for human processing, with
a few ideas in each knowledge block elaborated
in subordinate knowledge blocks. The
knowledge structure obtained by such successive elaborations, starting from some toplevel knowledge block, is then predominantly
hierarchical. (However, it has ultimately the
characteristics of a network because overlapping knowledge and cross-references connect
knowledge blocks in ways other than
superordinate-subordinate relation.)
In particular, the top-level knowledge block
describes the knowledge about the entire domain at the grossest level of description. The
lower-level knowledge blocks, obtained by successive elaboration, then provide progressively
more detailed descriptions of selected aspects
of knowledge described at higher levels. The
kinds of descriptive concepts and the symbolic
representations match the level of description.
Thus the higher-level knowledge blocks are
described in terms of abstract or vague concepts, expressed in the form of words or pic- '
tures, useful for unifying and subsuming more
detailed information. On the other hand, the
lower-level knowledge blocks are described
in terms of more precisely defined concepts
and are predominantly expressed in the
form of algebraic symbols and mathematical
formalism.

The knowledge base for any domain must


specify the particular entities of interest in this
domain,
the
particular
concepts
("descriptors") useful for describing these entities, the relations existing between such
descriptors, and procedures for transforming
such relations. In a quantitative science like
physics, most important descriptors are
"quantities" (i.e., descriptors characterized by
numerical values) and relations between such
descriptors are often expressed by
mathematical equations.
These essential features can be concretely exemplified in the specific domain of mechanics
(i.e., the science of morion). The top level of
the knowledge base for this domain describes,
at a gross level, the entire knowledge about
mechanics and is summarized in Figure 1. The
particular entities of interest in mechanics are
"particles" (i.e., objects small or simple
enough to be specified by the positions of
single points) and "systems" consisting of one
or more particles (e.g., the planets, a string,
etc.). The aim of the science of mechanics is
then to describe and predict how such systems
move in the course of time.
This aim is furthered by introducing two
different kinds of descriptors, "individual
descriptors" and "interaction descriptors".
An "individual descriptor" is a concept used
to describe a system without attention to the
interactions of its particles with each other or
with other particles. Thus an individual
descriptor of a particle involves only specification of this particle alone. (For example, individual descriptors of a particle i include intrinsic descriptors such as its ' 'mass' ' m\ or motion descriptors such as its "velocity" v.) By
contrast, an "interaction descriptor" is a concept used to describe how particles or systems
interact with each other. Thus an interaction
descriptor of a particle requires the specification of this particle and of the other particles
with which it interacts. (For example, the interaction of some particle / with some other
particle/may be described by the "force" F
exerted on i by/.) There are also composite interaction descriptors (e.g., "energy") which
involve combinations of simpler interaction
descriptors and individual descriptors.
The essential structure of the entire science
of mechanics, as summarized in Figure 1, consists then of the following four types of knowledge: (a) One type of knowledge concerns the

KNOWLEDGE STRUCTURE AND PROBLEM SOLVING

OVERVIEW

105

OF KNOWLEDGE ABOUT MECHANICS

(entities: particles and systems thereof)

<

individual descriptors

interaction descriptors

(intrinsic and motion)

and relations between them.

and relations between them

Downloaded by [New York University] at 12:18 23 April 2013

interaction laws
rel (interaction descriptor,
individual descriptors)

motion principles
rel (time-change of motion descriptors,
interaction descriptors)
Figure 1. Top level of the knowledge base of mechanics. The notation rel (A, B) indicates a
relation between A and B.
definitions of various individual descriptors
and the relations between such descriptors, (b)
Another type of knowledge concerns the
definitions of various interaction descriptors
and the relations between such descriptors, (c)
An essential task of mechanics is to specify how
the interaction between any two systems
observed in nature depends on the properties
of these systems. This knowledge is summarized by "interaction laws" which specify the
relations between interaction descriptors and
the individual descriptors of the interacting
systems (e.g., which specify how the force on a
particle by another depends on the intrinsic
properties and positions of the interacting
particles), (d) The individual descriptors of
primary interest in mechanics are "motion
descriptors" (such as position or velocity)
used to describe the motion of particles.

The crucially important knowledge, whereby


mechanics achieves its predictive power, is
then summarized by "motion principles"
which specify how the motion descriptors of a
system change with time because of the interaction of this system with other systems.
In addition to the preceding essential
knowledge, summarized in Figure 1, the
knowledge base of mechanics includes also
knowledge about the properties of various
special systems (such as strings, rigid bodies,
fluids, etc.). Indeed, many of these properties
can be derived by applying the interaction laws
and motion principles to these special systems.
Note that we have deliberately introduced
special concepts, such as "individual descriptors" and "interaction descriptors", which are
somewhat abstract and not quantitatively
defined. These descriptors are useful for

Downloaded by [New York University] at 12:18 23 April 2013

106

F. REIF AND JOAN I. HELLER

describing the most essential characteristics of


the systems studied in mechanics. As we shall
see, they also help to subsume a wide variety of
more detailed quantitative descriptors introduced in the further elaboration of the
knowledge base.
Despite its lack of detail, the top-level
knowledge in Figure 1 is sufficient to indicate
how the science of mechanics achieves its
predictive power. Any motion principle
specifies how the motion descriptors of a
system change with time as a result of interactions with other systems. But the interaction
laws specify how these interactions depend, in
turn, on the intrinsic descriptors and motion
descriptors of the interacting systems. By combining these relations to eliminate any further
explicit reference to interaction descriptors,
one can then deduce how the motion descriptors of a system change with time and thus can
predict the motion of this system.

Individual Descriptors

Elaboration of the
Knowledge Base
The following paragraphs outline how the
top-level of the knowledge base for mechanics,
illustrated in Figure 1, is elaborated down to
levels of detail where general relations are expressed in the form of precise mathematical
equations.

Figure 2 and Table 1 elaborate the knowledge about individual descriptors and the relations between them. Some of these descriptors
(e.g., "mass" or "electric charge") are "intrinsic descriptors" which describe inherent
characteristics of particles or systems. By contrast, the other individual descriptors (e.g.,
"position" or "velocity") are "motion
descriptors" which are defined by specifying
the position of a system by comparison with
some other system used as "reference frame".
(Many of these motion descriptors are "vectors", i.e., quantities specified jointly by both
a magnitude and a direction, and are indicated
symbolically by letters printed in bold-faced
type or underlined by a squiggly line.)
Table 1 indicates that the basic motion
descriptors of a particle are its ' 'position vector", its "velocity" (defined as rate of change
of position with time), and its "acceleration"
(defined as rate of change of velocity with
time). These definitions imply various special
relations between these descriptors in special
cases, such as for motion along a circle or for
motion along a straight line. (For example,
Table 1 includes information about the special
case of circular motion. The "components" of
a vector, mentioned there, are merely numbers
specifying the projections of the vector along
particular directions.)

(individual descriptors)

(intrinsic descriptors j

(particle)

(systeny)

(motion descriptors)

(particle)

( basic j
Figure 2. Individual descriptors and relations between them.

(system)

(complex)

KNOWLEDGE STRUCTURE AND PROBLEM SOLVING

107

Table 1
Elaboration of Motion Descriptors in Figure 2

Definitions are

Basic motion descriptors of a particle

indicated by =
position vector r
velocity v

v=

di/dt

acceleration a

a=

dv/dt

(speed v = magnitude
of velocity v)
(/ = time)

Special case: Motion along circle

Downloaded by [New York University] at 12:18 23 April 2013

rel(acce), speed, radius)

t ac = v2l r

(radial component of
a toward center)

l a = dv/dt

(tangential component
of a along velocity)

Complex motion descriptors of a particle

momentum p

p = wv

angular momentum 1

l =

kinetic energy K

K = y m v2

rxp

Motion descriptors of a system

momentum P

P =

angular momentum L

L = -II,
K = IK,

kinetic energy K
The "complex motion descriptors" of a particle (e.g., its "momentum", "angular
momentum", and "kinetic energy") are
defined as particular combinations of basic
motion descriptors and the intrinsic descriptor
"mass". They are introduced to express important motion principles in simplest possible
form. The motion descriptors ofa system (e.g.,
its "momentum" P, "angular momentum"
L, and "kinetic energy" K) are then simply
defined as sums of the corresponding complex
motion descriptors of all the particles in the
system. These are the motion descriptors used
to express the motion principles summarized
later in Table 4.
Interaction Descriptors
Figure 3 and Table 2 elaborate the knowledge about interaction descriptors and the

Ip,

relations between them. Different interaction


descriptors (e.g., "force", "torque", "potential energy") are introduced, again to express
important motion principles in simplest possible form. As shown in Figure 3 and more
specifically in Table 2, these different descriptors are interrelated by various defining relations. Furthermore, each descriptor can
describe interactions of various complexity,
e.g., the interaction between a particle and
another particle, or between a particle and a
system of particles, or between a system and
another system. As indicated in Figure 3 and
more specifically in Table 2, important principles specify how a descriptor, describing the
interaction between two complex systems, is
ultimately related to simpler descriptors (of
the same kind) describing the interactions between all die single particles in these
systems.

F. REIF AND JOAN I. HEUER

108

interaction descriptors
(force,torque;
work, potential energy, energy)

relation between any

relations between
different descriptors

descriptor and same

Downloaded by [New York University] at 12:18 23 April 2013

descriptor for interaction


between single particles
Figure 3. Interaction descriptors and relations between them.
Table 2
Elaboration of Interaction Descriptors in Figure 5

Relations between different descriptors

rel (torque, force)

for any particle

reKwork, force)

W = fFdr

rel (potential energy, work)

"A

WAS

(W = -At/)

for any particle;


for any path
for any system;
for work independent of process)
( WAS = work from state A
to standard state S.)

E=K+U

for any system


(K = kinetic energy)

rel (particle-particle force,


mutual particle-particle force)

F,-/ = - F / 7

"Newton's 3rd law"

rel (particle-system force,


particle-particle forces)

F/-Z/F

"superposition principle"

rel (system-system force,


particle-particle forces)

F-X*F

Tex, = total external force


(internal forces cancel)

rel (energy, potential energy)


Relations between same descriptors

For force:

* ext

KNOWLEDGE STRUCTURE AND PROBLEM SOLVING

viewed macroscopically). The interaction laws


for long-range interactions include those for
gravitational and electric interactions (e.g.,
Newton's famous law of universal gravitation
specifying the gravitational force between any
two particles). The interaction laws for most
short-range interactions are more empirical,
although ultimately derivable from more fundamental laws. For example, they include laws
describing the interaction between a particle
and a string, or between the touching surfaces
of two solid objects. Some of these laws are explicitly stated in Table 3.

Interaction Laws

Downloaded by [New York University] at 12:18 23 April 2013

Figure 4 and Table 3 elaborate the knowledge about interaction laws, i.e., they specify
how the interaction between two systems
depends on the properties of these systems for
various kinds of interactions. The interactions
occurring in nature can usefully be classified
into "long-range interactions" (which
manifest themselves even if the interacting
systems are separated by an appreciable
distance) and "short-range interactions"
(which become manifest only if the systems are
so close that they "touch" each other when

short-range interactions

long-range interactions

Kgure 4. Interaction laws.


Table 3
Elaboration of Interaction Laws in Figure 4
Gravitational interaction
Force
force on particle
by particle

FG

force on particle
by any system
(e.g., by earth)

F= mg

m\ 1712IR >

attractive

"Newton's gravity law"


(G = "gravitational
constant")
(g = "gravitational
acceleration")

Potential energy
potential energy
of two particles

U = Gm\m2i R

potential energy
of particle
near earth

U= m g h

String interaction
Force on particle
by string

109

F ^ 0, toward string
if string is taut
= 0 otherwise

if g = constant
(h = height up from
standard position)

110

F. REIF AND JOAN I. HELLER

Downloaded by [New York University] at 12:18 23 April 2013

Motion Principies
Figure 5 and Table 4 elaborate the centrally
important knowledge about motion principles. Some of these, like Newton's principle
about the acceleration of a particle (his ' 'second law of motion") arc basic principles which
provide a description sufficiently detailed to
predict the motion of every panicle in a
system. Other principles (like the momentum
or energy principles) can be derived from such
basic principles and provide a description of a
grosser kind. They are very useful because they
predict general motion properties without the
need for more detailed considerations. Note
that each of these principles specifies the rate
of change with time (didt) or the change with
time ( A ) of some quantity. In special cases
where certain interactions are absent, each of
these principles implies a corresponding ' 'conservation law" (e.g., the law of conservation of
energy) which asserts that some quantity (e.g.,
the energy) remains constant in time.

Functional Knowledge
Components
In the preceding sections the knowledge
base for mechanics has been elaborated down
to a level of detail where small clusters of important definitions or principles are expressed
in the form of precise mathematical relations.
These relations are the basic knowledge components of the science of mechanics and each
of them can be elaborated further. Indeed, if
such a knowledge component is to be functionally useful, it must be accompanied by
considerable ancillary knowledge.

detailed

Such ancillary knowledge includes


knowledge needed to interpret concepts and
relations (e.g., knowledge of procedures used
to define concepts operationally, knowledge
needed to achieve unambiguous specifications,
knowledge needed to make fine discriminations and to avoid confusions). It also includes
the knowledge of how to apply each relation to
solve routinely some primitive kinds of problems (e.g., how to find any quantity in a
relation from the others, how to find the
change in any quantity from a change in
another, and how to solve such problems
qualitatively or quantitatively in various symbolic representations).
The explication of the ancillary knowledge
needed to make a concept or relation functionally useful is worthy of study in its own
right. It also provides a theoretical basis for
predicting many of the "bugs" commonly exhibited by students. However, elaboration of
this ancillary knowledge would carry us too far
afield and is discussed elsewhere (Reif, Note
1). For the purposes of the present article we
shall thus merely presuppose, without further
analysis, the existence of the ancillary
knowledge needed to interpret and use each of
the concepts and relations listed in Tables 1
through 4.
Basic functional knowledge components,
each consisting of a definition or principle
together with the ancillary knowledge for its
effective use, provide a problem solver with
basic kinds of problems which are routinely
solvable ( i . e . , with basic " p r o b l e m
schemata"). In addition, the knowledge base
of more experienced problems solvers includes an accumulation of more complex functional knowledge components specifying how

description

Newton's
acceleration
principle
Figure 5. Motion principles.

momentum
principle

angular
momentum
principle

energy
principle

KNOWLEDGE STRUCTURE AND PROBLEM SOLVING

111

Table 4

Downloaded by [New York University] at 12:18 23 April 2013

Elaboration of Motion Principles in Figure 5


Name of Principle

Principle

Applicability

Output(O) and Advice(A)

Newton's acceleration
principle
("2nd law")
Momentum principle
(same as Newton's principle
if system is
single particle)

m a=F

inertial frame

0: reKaccel, interaction);
reKpos or vel, time)

d?ldt - F,

any particle
inertial frame

Angular momentum
principle

dL/dt = Tex,

Energy principle

AE = Wres

any system

inertial frame
or center-ofmass frame
any system
inertial frame
any system

commonly occurring problems of greater complexity can be solved routinely by the joint application of several principles.
Application Guidelines
The factual knowledge in the knowledge
base is accompanied by explicit "application
guidelines" to facilitate appropriate retrieval
and use of this knowledge for problem solving.
These guidelines include auxiliary factual and
procedural information of the following kinds.
Applicability Conditions
Concepts and principles in the knowledge
base are accompanied by "applicability conditions" specifying when these can legitimately
be used. For example, almost every motion
principle listed in Table 4 is accompanied by
the applicability condition specifying that the
principle is only valid if applied to a system
whose motion is described relative to special
kinds of reference frames called "inertial
frames" (e.g., relative to the fixed stars or approximately relative to the earth, but not
relative to any reference frame accelerated
relative to these). Similarly, the definition of
potential energy in Table 2 is accompanied by
the condition specifying that this concept is
only applicable in the case of special kinds of

0: reUvelocities,
external interaction);
A: choose system so that
wanted forces are external.
unwanted forces internal.
0: reHArotation velocities,
external interaction);
A: choose system so that
wanted torques are external.
unwanted torques internal.
0: reKAspeeds, Apositions),
no explicit mention of time.
( Wrn = residual work by interactions not included in )

interactions (those for which work is independent of the process whereby a system is brought
from one state to another).
Output Information
Important concepts or principles in the
knowledge base are accompanied by explicit
knowledge specifying the kinds of ' 'output information" obtained when such a concept or
principle is applied (sometimes together with
related knowledge). By comparing the output
information expected from a principle with the
goals pursued in a particular problem, better
decisions can be made about whether this principle could usefully be applied.
To illustrate how the availability of output
information can facilitate decisions in problem
solving, consider the following two simple
problems, A and B, both dealing with the motion of a baseball hit by a batter. Problem A
asks how much time elapses before the ball
reaches some specified height. Problem B asks
with what speed the ball moves when it reaches
this height. Newton's acceleration principle
and the energy principle are motion principles
applicable to both problems. But the output
information, accompanying these motion
principles in Table 4, specifies that the energy
principle, unlike Newton's principle, yields information relating changes of speeds and positions without any explicit mention of time.

112

F. REIF AND JOAN I. HEUER

Hence one is led to the decision that it would


be preferable to apply the energy principle to
problem B, but to apply Newton's principle to
problem A which requires a more detailed
analysis involving the time. (Although
Newton's principle could also be applied to
problem B, the solution would be more
laborious.)

Downloaded by [New York University] at 12:18 23 April 2013

Heuristic Rules
Factual information in the knowledge base
is often accompanied by heuristic rules which
provide advice about when and how to use this
information. Such advice can greatly facilitate
decisions in problem solving, although the advice may be overruled upon closer examination
of special circumstances. The following kinds
of heuristic rules deserve special mention.
There are description rules which specify
how a situation can usefully be described before particular knowledge is applied. For example, description rules can be formulated to
specify how situations in mechanics can usefully be described to facilitate the application of
the motion principles in Table 4. We shall
State these rules explicitly in the next section
when we discuss more fully the procedures
used to construct initial problem descriptions.
There are decision rules which provide advice about when particular knowledge might
usefully be applied. For example, the motion
principles in Table 4 are accompanied by advice that principles providing less detailed
descriptions (e.g., momentum or energy principles) should preferentially be applied before
principles providing more detailed descriptions (e.g., Newton's acceleration principle),
because they can often yield valuable information with relatively little effort, even if
they need later to be supplemented by more
detailed information.
Finally, there are implementation rules
which provide advice about how to use particular knowledge. For example, the momentum principle in Table 4 is accompanied by the
advice that it is best applied to a system
selected so that all forces, whose values are to
be found, are external forces; and so that all
forces, whose values are of no interest, are internal forces (because such internal forces are
not involved in the principle). As another example, the global knowledge of mechanics, illustrated in Figure 1, is accompanied by the
following advisory rules for applying this

knowledge: (a) Start with a motion principle


selected from the knowledge about motior
principles, (b) Relate the motion descriptors in
this principle to more primitive motion
descriptors (of greater interest in the particular
problem) by using the knowledge about individual descriptors and relations between
them, (c) Relate the interaction descriptors in
the motion principle to more primitive interaction descriptors by using the knowledge about
interaction descriptors and relations between
them. Then relate these more primitive interaction descriptors to individual descriptors by
using the knowledge about interaction laws.

Domain of Applicability
of the Knowledge Base
The knowledge base, summarized in Figures
1 through 5 and Tables 1 through 4, is (except
for some lower-level knowledge omitted for
the sake of brevity) sufficient to solve all the
mechanics problems commonly encountered
in an introductory college-level physics course.
An important property of any knowledge
base is its generalizability, i.e., its ability to
encompass more extensive knowledge without
major restructuring. In the case of mechanics,
the knowledge base presented in the preceding
pages has this desirable property. Thus the
basic structure of the knowledge base would
remain unaffected even if it were greatly
enlarged to encompass knowledge of classical
mechanics at the advanced graduate level. In
this case, Figure 5 and Table 4 would include
some additional motion principles yielding
detailed descriptions (e.g., "Lagrange's equations" and "Hamilton's equations"), where
these principles would be expressed in terms of
additional interaction descriptors included in
Figure 3 and Table 2 (e.g., "Lagrangian" and
"Hamiltonian" functions). But the top levels
of the knowledge base, and the essential structure summarized in Figures 1 through 5,
would remain unchanged.
GENERAL PROCEDURES
As mentioned in the introduction, the
knowledge base about a particular scientific
domain is to be used in conjunction with
general procedures whereby this knowledge
can be judiciously used to find the solutions of
diverse problems. The present section discusses

KNOWLEDGE STRUCTURE AND PROBLEM SOLVING

such general procedures facilitating problem


solving in physics, with particular emphasis on
the prototype domain of mechanics.
In particular, we shall discuss the following:
(a) procedures for initially describing and
analyzing any problem, with the aid of the
knowledge base, to bring it into a form
facilitating its subsequent solution, (b)
methods used to aid in the search and decision
processes needed to construct the actual solution of the problem, (c) ways of assessing the
resulting solution to test whether it is correct
and optimal so that appropriate revisions can
be made.

any process specified in the problem, and the


expression of the information in several useful
symbolic forms (including pictorial representations particularly useful for describing visually
observable phenomena). The following major
steps outline a procedure guiding a human
problem solver to construct a basic description
of any problem:

GENERATION OF A BASIC DESCRIPTION

Downloaded by [New York University] at 12:18 23 April 2013

Problem Description and Analysis


The initial description of a problem can
crucially determine how easily the problem can
be solved or whether it can be solved at all. In
the words of Charles Kettering, " a problem
well stated is a problem half solved."
Hence the first task faced by a problem
solver is to redescribc an original problem in a
way facilitating the subsequent search for its
solution. This can be done by identifying the
knowledge relevant to the problem, organizing it effectively, and describing it in convenient symbolic form. The resulting problem
description can thereby limit and subdivide
the domain of search and put the problem in a
form permitting ready application of the
problem-solver's knowledge base.
The construction of a useful problem
description can be separated into two fairly
distinct stages. The first stage uses
predominantly everyday knowledge to
generate a "basic" problem description readily interpretable by the problem solver. The
second stage then exploits fully the problemsolver's specialized knowledge about the relevant domain to generate a "theoretical"
problem description. These two kinds of descriptions will now be discussed in turn.
Basic Description
The aim of the basic description is merely to
translate an originally presented problem into
a form describing clearly the situation specified
in the problem and the information to be
found. In particular, this requires the identification of specified systems and their properties, a detailed time-sequenced description of

113

Specified situation
Use diagrams and/or statements to describe the
specified information about each system and its
properties. Introduce convenient symbols and
identify those denoting unknown values
(wanted or ultimately unwanted).
Constants: Summarize the time-independent features.
Process;
If some features arc time-dependent, envision the corresponding process by describing the situation at successive times.
Identify any subprocesses in which the description is distinctly different. Describe the
situation at a typical time and at the extreme
times of each such subproccss.

Goal
Summarize the wanted information in terms of
previously introduced concepts and symbols.

To exemplify such a basic problem description, consider the following sample problem
described in the original form in which it
might typically be presented to a human
subject:
PENDULUM PROBLEM
A pendulum consists of a small ball attached to one
end of a light string of length L. The other end of
the string is attached to a hook fastened to the ceiling. A fixed peg is located vertically below the hook
at a distance smaller than L. The ball is initially
held at rest, with the string taut and horizontal,
and is then released. What must be the minimum
distance between the hook and the peg so that the
string is still taut when the ball reaches a point
directly above the peg?

When the previously outlined procedure is


implemented, this problem statement is
translated into the basic description summarized by Figure 6 and the following statements:
(Note that the process specified in the problem
has been explicitly described and decomposed
into the two subprocesses from time / to t\
and from /j to t^. The symbol /j denotes a
typical time between t\ and ty)

114

F. REIF AND JOAN I. HELLER

ball

hook

'1,2
Known: L

Wanted: D

min
(Other parameters unknown and ultimately unwanted)
Downloaded by [New York University] at 12:18 23 April 2013

Figure 6. Diagram summarizing the basic description of the pendulum problem stated in the text.
BASIC DESCRIPTION OF
PENDULUM PROBLEM
(Sec also Figure 6)
Constants:
String: length L.
Peg: distance D below hook, D<L.
Process:
Q:

String: taut and horizontal.


Bill: distance I to left of hook, at rest.

'O, 1: String: taut, angle 6 below horizontal.


Ball: moves down along circle of
radius L.
-ty.

String: taut and vertical, touches peg.


Ball: distance L below hook, moves to
right.

t\

2- String: taut, touches peg.


Ball: moves up along circle of radius
L-D.
Ball: distance L-D above peg, moves to
left.

Goal:
'min= ? (if string is taut at time t-j)

Although a basic problem description can


usually be constructed fairly easily by a human
problem solver, it is far from trivial: (a) Care
must be taken to avoid common mistakes,
e.g., either ignoring some information
specified in the original problem or making
some unwarranted assumptions, (b) Commonsense knowledge must be used to explicate
default conditions implied in the problem.

For example, in the absence of more specific


information, the pendulum problem has been
interpreted by assuming that the ceiling is that
of a room on the earth (rather than that of a
traveling spaceship), that the string does not
stretch appreciably so that its length remains
constant, and that the sizes of the ball and peg
are negligibly small compared to the lengths L
and D. (c) Primitive or qualitative knowledge
is also needed to describe more fully processes
partially specified in the problem (e.g.,
familiar knowledge about the motion of objects is used to assume that the released ball
starts moving down rather than up).
The construction of a basic description translates a problem into a more transparent form
facilitating its solution or further redescription. Indeed, the next section discusses specifically how a problem can be further simplified
by being redescribed in terms of specialized
theoretical knowledge. In particular, such a
theoretical redescription provides important
checks on interpretations and assumptions
made in the previous basic description.
Theoretical Description
The knowledge base about a particular
scientific domain specifies the entities of interest in this domain, the special concepts
useful for describing these entities, and important principles relating these concepts. If a
problem is deliberately redescribed in terms of
these concepts, the entire theoretical
knowledge about this domain becomes then
readily accessible to implement the solution of
the problem.

KNOWLEDGE STRUCTURE AND PROBLEM SOLVING

The generation of such a "theoretical


description" is facilitated if the knowledge
base about a domain contains explicit heuristic
rules (of the kind mentioned in the preceding
section) which specify how to describe effectively any situation encountered in this domain. The following are the most essential
steps of such description rules for the domain
of mechanics:

Downloaded by [New York University] at 12:18 23 April 2013

Generation of a Theoretical Description


Relevant times and systems: At each relevant
time (e.g., typical and extreme times of previously identified subprocesses) identify those
systems, of interest in the domain of mechanics,
which are relevant in the problem because information about them is wanted, or because they
interact with such systems directly or indirectly.
Description of relevant systems: At each relevant
time, describe each relevant system as follows,
introducing convenient symbols and expressing
simply related quantities in terms of the same
symbol:
Description of motion: Draw a "motion
diagram" indicating all available information
about the position, velocity, and acceleration
of each particle relative to some specified reference frame.
Description of interaction: Draw a "force
diagram" indicating all external forces on the
system. Identify these forces as follows:
Long-range forces: Identify all other objects interacting with the given system by
long-range interactions (i.e., ordinarily just
the earth interacting by gravitational interaction). For each such interaction indicate
on the diagram the corresponding force and
all available information about it.
Short-range forces: Identify every other object which touches the given system and
thus interacts with it by short-range interaction. For each such interaction indicate
on the diagram the corresponding force and
all available information about it.
Checks of description: Check that the descriptions of motion and interaction are qualitatively
consistent with known motion principles (e.g.,
that the acceleration of each panicle has the
same direction as the total force on it, as required
by Newton's motion principle mi = F).

Note that the preceding description procedure exploits extensively the specialized information in the knowledge base of
mechanics. Thus it describes explicitly every
system in terms of the special motion descriptors (e.g., velocity and acceleration) and interaction descriptors (forces) specified by the
knowledge base. It deliberately classifies

115

interactions into long-range and short-range


types to formulate explicit criteria for identifying all forces on a system. It incorporates in the
description the known relations of Table 1
about motion descriptors (e.g., relations between acceleration and velocity for circular motion), the known relations between interaction
descriptors in Table 2 (e.g., Newton's third
law specifying that the magnitudes of the
mutual forces between interacting particles are
equal), and the known interaction laws of
Table 3 (e.g., the known properties of gravitational forces or forces exerted by strings). Finally, it uses the motion principles of Table 4 to
check that the resulting theoretical description
is internally consistent.
Figure 7 illustrates the theoretical description generated by the preceding description
procedure for the pendulum problem whose
basic description was previously summarized in
Figure 6. At each of several times, Figure 7 describes the motion of the ball (relative to the
ceiling) by information about its velocity v and
acceleration a. At these times Figure 7 also
describes the interaction of the ball by information about the gravitational force Fg on the
ball by the earth and the force Fs on the ball by
the string.
In the example of Figure 7, the theoretical
knowledge of Table 3 has been used to assert
that the gravitational force Fg is downward and
equal to mg (where m is the mass of the ball
and g is the gravitational acceleartion near the
earth), that the force Fs exerted by the string is
directed along the string away from the ball,
and that the magnitude of this force must be
larger than zero if the string is taut. (In the
absence of more specific information, default
conditions have been invoked to neglect any
frictional forces by the surrounding air.) Finally, the condition that the direction of the acceleration must always be the same as that of
the total force (as required by the motion principle m = F) has been used as a check on the
assumed description of the entire motion process. In particular, it has been used to infer
that, at the time ti when all forces on the ball
arc downward, its acceleration must also be
downward.

Simplifications Introduced
by Problem Description
The initial description of a problem can
greatly facilitate the subsequent search for its

116

F. REIF AND JOAN I. HELLER

Downloaded by [New York University] at 12:18 23 April 2013

'0,1

'1,2

Figure 7. Theoretical description of the pendulum problem described in Figure 6.


solution, sometimes to the extent that the
solution becomes obvious. The following examples illustrate the kinds of simplifications
introduced by the initial description.
To illustrate simplifications introduced by
the description of interaction, consider again
the pendulum problem of Figure 6. In this
case the theoretical description of interactions,
shown in Figure 7, translates the original condition that "the string is still taut at the time
t2" into the equivalent condition that "the
force exerted by the string is still non-zero at
the time t2". As a similar example, consider a
problem which asks when a sled, sliding down
along the snow-covered surface of a mountain,
flies off this surface. The theoretical description of interactions translates this question into
one asking when the force exerted on the sled
by the surface becomes zero. In both of these
examples, the redescription of the problem in
terms o forces removes a major difficulty of interpretation and makes the problem readily
solvable by familiar motion principles involving forces.
To illustrate simplifications introduced by
the description of motion, consider the problem, shown in Figure 8, where two particles A
and B are attached to a string passing over a
fixed pulley. The question is to find the velocity with which the particle A strikes the wall

after the particles are released from rest from


their initial positions in Figure 8. The main
difficulty of this problem lies in the description of the motion. Thus the basic description
requires envisioning how the particles move.
Then the theoretical description requires taking into account the fixed length of the string
to relate the positions or velocities of the two
particles at any time. After this initial description has been accomplished, the solution of
the problem can be easily found by a simple
application of the energy principle.

pulley

L/2

*B
Figure 8. Diagram illustrating a problem
about two particles connected by a string passing over a pulley.

Downloaded by [New York University] at 12:18 23 April 2013

KNOWLEDGE STRUCTURE AND PROBLEM SOLVING

117

Qualitative Analysis

Constraint Satisfaction

The theoretical description of a problem is


usefully followed by a qualitative analysis of
the problem. Such an analysis explores
qualitatively the main implications of some
principles applicable to the problem. (For
example, in the case of the pendulum problem, the energy principle leads to the
qualitative implication that the speed of the
ball increases when it descends along its
circular path, and then decreases when it
ascends along its new circular path). The
qualitative analysis also explores the implications expected for different possible values of
the parameters in the problem, particularly for
extreme values. (For example, if the distance D
is almost as large as the length L in the pendulum problem, the string would not be expected to become slack; but if D is much
smaller than L, the string would be expected to
become slack).
Space limitations prevent us from discussing
the qualitative analysis more fully. Suffice it to
say that such an analysis may facilitate the
subsequent solution of a problem by suggesting possible approaches, by helping to
avoid errors and pitfalls, and by helping to interpret physically the results of mathematical
arguments.

A constraint is a formal statement expressing


a special condition relating the values of the
descriptors in a particular problem. In its most
detailed form, such a constraint is commonly
expressed in the form of an equation or inequality relating the numerical values of various
parameters. For example, in the case of the
pendulum problem illustrated in Figures 6 and
7, the following three constraints relate the
lengths L and D, the magnitude v2 of the
velocity of the ball at the time t2, and the
magnitude a2 of its acceleration at this time:

Construction of Solution
The initial description and analysis of a
problem pave the way for the actual construction of its solution. The search for the solution
is greatly facilitated by some generally applicable methods used in conjunction with the
knowledge base about a particular scientific
domain. Hence we devote this section to
discuss and explicate such general methods
useful in mechanics as well as throughout all of
physics. In particular, we shall discuss powerful
methods of constraint satisfaction, used
together with systematic ways of decomposing
problems and exploring decisions. We shall
also examine the use of multiple levels of
description for exploring decisions and planning. Finally, we shall indicate how the search
for a solution is facilitated by the joint
application of all these methods, when used
together with a well-structured knowledge
base and a well-formulated prior description of
a problem.

Vimv22 + mg (L-2D) = 0.
The methods used to generate such constraints
will be the main topic discussed in the next few
sections.
Every constraint narrows the range of possible
solutions consistent with it. This remark is the
basis of a powerful method of "constraint
satisfaction" which is used extensively
throughout all of physics and which consists of
the following two essential steps: (a) Generate
enough constraints so that only one solution exists that is consistent with all of them, (b) Construct an actual solution that satisfies all these
constraints. This must then be the solution.
The second step, i.e., the construction of a
solution from available constraints, is usually fairly straightforward (unless one deals with problems far beyond the level of basic college-level
physics). Indeed, such a solution can be constructed by one or both of the following techniques: (a) Formal mathematical techniques exist
for modifying or combining existing constraints
to obtain more useful constraints and thus
ultimately to obtain the desired solution. (A
well-known example is provided by the standard
algebraic techniques for solving simultaneous
algebraic equations with several unknown
variables. For instance, once one has generated
the three previous constraints for the pendulum
problem, such algebraic techniques allow one to
find the minimum value of D immediately and
thus to solve the problem.) (b) Systematic techniques of generating and testing can be used with
great effectiveness. These techniques involve judiciously guessing xheform of the solution, often
with some adjustable parameters, then checking
whether this assumed solution can be made to
satisfy all constraints, in which case it must be the
solution. (For example, such techniques arc commonly used to solve differential equations.)

F. REIF AND JOAN I. HELLER

Downloaded by [New York University] at 12:18 23 April 2013

118

The method of constraint satisfaction has


the following advantages: It can easily be used
to solve complex problems which cannot be
solved by search methods that proceed more
sequentially between the specified information
and the goal of a problem. (Such sequential
methods, or cause-and-effect arguments, can
be viewed as special cases of constraint satisfaction.) Constraints can be generated rather uncritically, in any order, and may later be
transformed into more desirable constraints.
Furthermore, even if generation of a valid constraint does not help convergence toward the
desired solution, it can never cause divergence
from it.
To exploit fully these advantages of constraint satisfaction, one must have efficient
methods for making the decisions needed to
generate and transform constraints. We shall
discuss such methods presently.
Problem Decomposition
in Constraint Satisfaction
To make judicious decisions about what
constraints to generate and how to transform
them, problems are decomposed into successive subproblems by a heuristic method involving repeated application of the following
major steps:
(1) Description of current problem: Identify currently known and desired information
and organize this information according to
estimated importance.
(2) Choice of subproblem: Identify some
difficulties to the solution of the problem.
Then select a particular problem aiming to
reduce one of these difficulties.

(3) Choice of method: Identify some promising methods for solving the chosen subproblem. Then select a particular useful
method.
(4) Application of method: Apply the
chosen method and describe the result.
The following comments are designed to explain these steps somewhat more fully.
Description of current problem. The
description of a currently considered problem
involves, in particular, an identification of
those parameters whose values are currently
known, those whose values are unknown and
wanted to attain the problem goal, and those
auxiliary parameters whose values are

unknown and ultimately unwanted. In addition, the description seeks to facilitate decisions by focusing attention selectively on those
constraints most important for further progress
toward the solution. Hence the available information is categorized so that information
which has already been translated into contraints, or constraints which have already been
transformed into more useful constraints, are
classified as no longer important.
Problem difficulties and subproblems. As
summarized in Table 5, the following kinds of
difficulties, in order of decreasing severity,
may present obstacles to the solution of a problem: (a) There may be a lack of sufficient constraints, needed to solve the problem, because
some relevant knowledge about the problem
has not yet been translated into such contraints, (b) The available constraints may contain parameters which are ultimately unwanted, (c) The available constraints may have
an unwanted form (e.g., such a constraint may
involve x2 rather than x itself).

Table 5
Problem Difficulties and Available Methods
Difficulties

Remedial subproblem

Methods

lacking constraints

generate particular
kind of constraint

instantiate general knowledge


or
explicate particular knowledge

constraints with
unwanted parameters

remove particular
unwanted parameters

combine constraints
or
generate needed constraint

constraints with
unwanted form

remove particular
unwanted form

modify constraint

Downloaded by [New York University] at 12:18 23 April 2013

KNOWLEDGE STRUCTURE AND PROBLEM SOLVING

According to a simple means-ends analysis,


each such difficulty suggests that it may be
useful to solve a corresponding subproblem
aiming to remove this difficulty. Potentially
useful subproblems involve then either (a)
generating a particular kind of constraint, or
(b) removing some particular unwanted
parameters, or (c) removing some particular
unwanted form. It is then necessary to choose
selectively one of these subproblems for consideration.
Available methods. Table 5 summarizes the
available methods for dealing with each of the
subproblems just mentioned. These methods
deserve fuller discussion since they reveal explicitly the most important decisions needed to
solve problems by the method of constraint
satisfaction.
The generation of a new constraint is the
most complex of the relevant subproblems.
Sometimes such a constraint may be generated
by merely expressing, in explicit mathematical
form, some information specified in the problem itself. But much more commonly a constraint 'is generated by "knowledge instantiation", i.e., by applying general knowledge
from the knowledge base to particular information contained in the problem. Specifically,
such knowledge instantiation involves,
"Applying <some general principle or
definition > to <some particular entity > at
<some particular time or during some particular time interval > with < some particular
description > . "
Each of the preceding angular brackets indicates explicitly some component choice
which must be made in the complex decision
needed to generate a constraint. To elaborate
somewhat further, the choice of a particular
entity often involves choosing between some
primitive entity or a composite entity consisting of several primitive entities (e.g.,
deciding whether to apply a motion principle
to a link in a chain or to the entire chain).
Similarly, the choice of a time interval often
involves choosing between some time interval
or a more inclusive composite time interval
consisting of several shorter time intervals. The
choice of description commonly involves decisions such as the following: Relative to which
reference frame should one choose to describe
the motion of particles? Relative to which
direction should one choose to describe the
components of vectors? Relative to which standard position should one choose to describe
the potential energy of particles?

119

To illustrate the generation of constraints by


knowledge instantiation, consider again the
pendulum problem described in Figures 6 and
7. In the solution ofthat problem (outlined in
the Appendix) one important constraint is
generated by, "Applying <Newton's acceleration principle > to < the ball > at < the
time t2> with <motion described relative to
the ceiling > and with < components chosen
along the downward direction>." Another
important constraint is generated by, "Applying <the energy principle > to <the ball>
between <the times t0 and t2> with <motion described relative to the ceiling>."
Removing unwanted parameters from some
constraints is a subproblem which can be easily
solved by using one or both of the following
methods: (a) One can use standard algebraic
techniques to combine existing constraints to
generate new constraints which involve none of
the unwanted parameters. As many as N unwanted parameters can thus be removed if
there are (N +1) available independent constraints involving these parameters, (b) If all of
the unwanted parameters cannot be removed
in this way because there are too few available
constraints, one can generate new constraints
involving those parameters that cannot be
removed. The generation of these constraints,
by the methods discussed previously, paves the
way for the subsequent removal of these
parameters by combining constraints.
Removing an unwanted form of a constraint
is a subproblem which can be solved by simple
algebraic techniques for modifying the form of
a constraint. Such techniques (e.g., performing the same operation on both sides of an
equation) are too routine to merit further
discussion in the present context.
It is apparent that additional decisions must
be made to decide which particular subproblem to consider, which particular method
to choose, and what choices to make in applying this method (particularly the complex
choices needed in knowledge instantiation).
Such decisions are facilitated by procedures
discussed in the next paragraphs.
Exploration of Decisions
A decision among a few available alternatives can be made judiciously by "exploring
the decision", i.e., by a procedure whereby
choices are made on the basis of an assessment of anticipated consequences. Such a

Downloaded by [New York University] at 12:18 23 April 2013

120

F. REIF AND JOAN I. HELLER

procedure requires more processing, but fewer


demands on memory, than one based on
stored rules prescribing choices to be made
under many possible conditions. It also allows
human problem solvers to make flexible
choices in diverse and unfamiliar situations.
To explore a decision one must be able to
characterize any alternative course of action by
its "utility", i.e., by a rough measure of its
benefits compared to its costs. The "benefits"
characterize the state resulting from the action.
They are considered positive or advantageous if
the resulting state is closer to the desired problem goal or leads to a simplification of the
problem; they are considered negative if the
opposite is the case. By contrast, the "costs"
characterize the process of carrying out the action and are considered large if this process requires much effort or time.
The utility, which is a rough measure of the
ratio of such benefits compared to such costs,
needs only be accurate enough to permit an
approximate ranking of alternatives according
to their estimated utility, i.e., estimated on
the basis of currently available information.
The estimated utility of a particular alternative
may thus change as more information becomes
available during the problem-solving process.
(Because the utility involves a comparison of
benefits and costs, a course of action offering
large benefits may nevertheless have zero utility if it cannot be implemented by the problem
solver. Conversely, an alternative offering only
minor benefits may have a large utility if it can
be implemented very easily.)
On the basis of the preceding comments,
the procedure for exploring a decision may be
summarized by the following major steps:
1. Identification of promising alternatives:
Identify, on the basis of available knowledge,
one or more alternatives of comparatively large
estimated utility.
2. Selection by exploration: Repeat the
following until the estimated benefits of finding a more useful alternative outweigh the
costs of further exploration: (a) Explore an
identified alternative by predicting its major
consequences and then assessing its estimated
utility on the basis of this information, (b)
Select, among the alternatives explored, that
of largest estimated utility.
3. Revision of unsatisfactory choices: (a) Implement the selected alternative, (b) If its actual utility is unsatisfactorily small, backtrack
and reconsider the decision process on the basis
of the new information now available.

Note that this decision process involves progressively more refined choices made on the
basis of increasingly more reliable information.
Thus the first step makes preliminary choices on
the basis of general information stored in the
knowledge base (e.g., on the basis of the
previously discussed "application guidelines"
included in the knowledge base); the second
step makes improved choices on the basis of information about anticipated consequences; and
the third step may revise such choices on the
basis of information about the actual consequences of an implemented action.
Note also that the second step of the decision
procedure may consider only the most obvious
alternative alone if it seems sufficiently useful.
Thus the decision procedure does not aim to
make an optimum choice, but merely one
which is good enough (i.e., which "satisfices").
Despite its seeming complexity, the decision
procedure just outlined explicates to some extent processes which persons perform quite
commonly in daily life. To provide an example
of the exploration of decisions in physics problems, consider a typical kind of decision
encountered in an attempt to generate constraints by "knowledge instantiation". In the
specific case of the pendulum problem described in Figures 6 and 7, the application
guidelines in the knowledge base suggest that it
would be useful to apply the energy principle to
the ball. However, a further choice must still be
made about the times between which this principle should be applied. One alternative, suggested by the problem description in Figure 7,
is to apply this principle twice once between
the times t, and t\ (when the ball is at its lowest
point) and then again between the times t\ and
2- Another alternative is to apply this principle
during the entire time interval between the
times t0 and 2- The first alternative would lead
to the generation of an extra intermediate result
(an additional constraint involving the unwanted velocity at the time ti) and would thus
involve more processing to obtain a solution.
Hence the second alternative has larger estimated utility and would be chosen preferentially.

Multiple Levels of Description


A relation expressing a constraint can be
described at various levels of detail. The most
detailed description is in the form of a
mathematical equation which specifies precisely what parameters are related and how they

Downloaded by [New York University] at 12:18 23 April 2013

KNOWLEDGE STRUCTURE AND PROBLEM SOLVING

are related. [For example, Newton's motion


principle ma. = F involves three parameters
(the mass wz of a particle, its acceleration a,
and the force F on it by all other particles) and
specifies precisely how these parameters are
related.] A less detailed description might
merely specify what parameters are related
without specifying how they are related [e.g.,
Newton's principle might merely be described
by: rel(*zz,a,F), i.e., as some relation, of
unspecified form, between the indicated
parameters.] An even less detailed description
might omit mention of some of the parameters
and might describe other parameters collectively in terms of more abstract or qualitative
concepts. [For example, Newton's principle
might thus be simply summarized as: rel(motion, interaction), i.e., as a relation between
some motion parameter and some interaction
parameter.]
The flexibility of using multiple levels of
description can be effectively exploited to
facilitate decisions in problem solving. (Accordingly, it is also helpful if the knowledge base
about a domain is desaibed at various levels of
detail.) For example, gross qualitative descriptions are very useful when exploring decisions,
in the manner discussed in the preceding section, to describe approximately the major anticipated results of contemplated actions and to
estimate their utility without becoming
enmeshed in complicating details.
Multiple levels of description are particularly
useful for planning solutions. Indeed, a problem can be solved at any level of description by
using the search methods discussed in the
preceding sections. Thus one can first construct
a schematic problem solution at a gross level of
description where decisions can be made
without attention to complicating details. This
solution can then be used as a plan guiding the
construction of a solution at a more detailed
level of description. This procedure may be
repeated successively until one obtains the
completely detailed solution.
Planning by successive refinements can
greatly facilitate search and provides advantages exploited in many fields (e.g., Hughes &
Michton, 1977; Sacerdoti, 1977). The essential
advantage is that any decision made at a gross
ievel of description need deal with only limited
information, unencumbered by excessive
details, but narrows greatly the range of alternatives to be considered subsequently in decisions at a more detailed level of description. In
this way the domain of search can be rapidly

narrowed down, although only few alternatives


are considered at each step. Note that these
essential advantages are preserved even if the
early gross descriptions of the problem are
vague or ambiguous.
For example, the following is a grossly
described solution of the pendulum problem
illustrated in Figures 6 and 7: (a) Apply a motion principle to the ball at the time t2 when its
interaction is known to be such that the string
is still taut. The result should be a relation
specifying a constraint on the motion of the
ball at the time t2. (b) Apply motion principles
to the ball between the times t0 and t2. The
result should be a relation between the motion
of the ball at the time t2 and its specified motion at prior times, (c) Combine the preceding
information. [Although this schematic solution is vague, it suffices to provide a plan
decomposing the solution into its two major
steps (a) and (b).]

Efficacy of Methods Used Jointly


with Knowledge Base
The methods discussed in the preceding sections are designed to help make judicious decisions facilitating the solution of physics problems. As the following comments indicate,
these methods are quite powerful if they are
used jointly and in conjunction with a wellstructured knowledge base.
Decisions about how to transform constraints can be readily made by using the
previously discussed heuristic procedures for
decomposing problems. The most difficult
decisions arc those needed to generate constraints (i.e., decisions about what principle to
apply, to what system, at what time, and with
what description). But these complex decisions
are substantially facilitated because the wellstructured knowledge base and prior problem
description greatly restrict the range of alternatives to be considered. The following
remarks illustrate this statement.
As discussed previously, the knowledge base
is hierarchically organized into knowledge
categories and subcategories which are accompanied by explicit application guidelines. In
particular, these guidelines suggest that the
solution of a problem start by application of
one of the motion principles. There arc only
four such principles in the knowledge base illustrated in Table 4. Furthermore, the application guidelines accompanying these principles

Downloaded by [New York University] at 12:18 23 April 2013

122

F. REIF AND JOAN I. HELLER

provide advice which ordinarily limits even


further the,number of principles useful in a
given situation. Decisions between these very
few principles can then easily be made after a
little exploration.
Once a selected motion principle has been
applied, the application guidelines in the
knowledge base provide guidelines about a
whole set of other decisions needed to
elaborate this motion principle. Indeed, as
already mentioned, they suggest that motion
descriptors in the principle should be related
to simpler motion descriptors (by using the
knowledge about motion descriptors in Table
1). They also suggest that interaction descriptors should be related to simpler interaction
descriptors (by using the knowledge about interaction descriptors in Table 2) and that these
should then, in turn, be related to individual
descriptors (by using the knowledge about interaction laws in Table 3). (For example, as
discussed more fully in the Appendix, one step
in the solution of the pendulum problem involves application of Newton's motion principle ma = f to the ball at the time t2. The acceleration a is then elaborated by relating it to
the velocity of the ball at this time. Similarly
the total force F on the ball is elaborated by
relating it to the gravitational force on the ball
by the earth and the force on the ball by the
string; each of these forces is then, in turn,
related to the individual properties of the ball
and of the string.)
The initial description of a problem helps to
identify what particular systems and times
should be considered in applying principles
from th knowledge base. (For example, the
descriptions of the pendulum problem in
Figures 6 and 7 identify the ball as the major
system of interest and identify the particularly
interesting times t0, tlt and t2 during its motion). Furthermore, the initial theoretical
problem description, by deliberately describing a situation in terms of knowledge provided
by the knowledge base, explicates much of the
knowledge subsequently needed to apply principles from the knowledge base and also helps
to make decisions about what descriptions to
use when these principles are applied. (For example, the description of the pendulum problem in Figure 7 shows that the acceleration of
the ball, and all forces on it, arc directed
downward it the time t2. It is then obviously
best to apply the motion principle /wa = F to
the ball by describing the principle in terms of
components along the downward direction.)

The preceding comments, which indicate


how solution methods and the knowledge base
interact to facilitate decisions in problem solving, are exemplified further in the Appendix
which outlines the detailed solution of the pendulum problem described in Figures 6 and 7.
Assessment of Solution
After a solution has been constructed, it is
important to assess how satisfactory it actually
is. We merely mention briefly the following increasingly stringent criteria which must be satisfied to ensure that a solution is free from obvious defects. Any constructed solution is tested
against these criteria. Failure to satisfy any one
of these criteria requires that the solution be
revised until all these criteria are satisfied.
Clear interpretation: The answers to a problem should only involve clearly defined
parameters and should specify all parameters
needed to avoid ambiguity (e.g., units of
measurement, magnitudes and directions of
vectors, reference frame used to specify motion
descriptors, etc.).
Completeness: (a) All questions should be
answered, (b) Every answer should be expressed entirely in terms of known quantities,
(c) Answers should be given for all types of
solutions possible for different values of the
parameters in the problem.
Internal consistency: (a) All steps in the
argument should be free of logical errors, (b)
Answers should be internally consistent (e.g.,
the units of quantities in an answer should be
consistent).
External consistency: (a) Answers should
agree with qualitatively expected functional
relationships between parameters in the problem, (b) Answers should agree with those obtained for extreme values and other special
values of the parameters in the problem.
Optimality: (a) The solution of the problem
should be as simple as possible, (b) The
answers should be expressed in the simplest
and/or most easily interpretable way.
DISCUSSION AND IMPLICATIONS
Scope and Limitations
In the preceding pages we sought to specify
important cognitive mechanisms facilitating
effective human problem solving in the

Downloaded by [New York University] at 12:18 23 April 2013

KNOWLEDGE STRUCTURE AND PROBLEM SOLVING

domain of physics. In particular, we first


discussed the content, organization, and
description of the knowledge base facilitating
such scientific problem solving. Then we
discussed procedures used to describe and
analyze a problem initially, to search for its
solution, and to assess this solution. Finally,
we pointed out that effective problem solving
depends crucially on the joint application of
such a knowledge base used in conjunction
with such procedures.
Since our main interest has been in human
scientific problem solving, our discussion has
explicitly presupposed basic informationprocessing capabilities characterizing human
subjects. However, some of the ideas discussed
are probably also germane to scientific problem solving by computers programmed by the
methods of artificial intelligence.
Although our discussion has dealt
predominantly with the domain of basic
mechanics, it could readily be extended to
other fields of physics and even to other quantitative or engineering fields. The particular
content of the knowledge base would then, of
course, be different. But its general
characteristics, (e.g., its hierarchical organization or its inclusion of explicit application
guidelines) would be similar. Furthermore, the
problem-solving procedures discussed by us are
quite generally applicable, although they need
to be used in conjunction with the particular
knowledge base for the scientific domain of
interest.
Space limitations did not permit us to
elaborate or exemplify several issues in greater
detail. Moreover, we omitted entirely the examination of some important topics. One such
topic concerns the knowledge of individual
scientific definitions and principles,
knowledge which we explicitly assumed to
preexist in the knowledge base. Such
knowledge is clearly a necessary, although not
sufficient, prerequisite for scientific problem
solving. Furthermore, the underlying
knowledge required to interpret and use any
particular definition or principle is considerable and merits detailed analysis (Reif,
Note 1). Deficiencies in such knowledge lead
also to many of the commonly observed errors
and misconceptions exhibited by students. A
second omitted topic concerns humanly useful
'bookkeeping procedures" for managing effectively complex information of the kind encountered during problem solving. Although
such procedures are fairly routine, they are

123

often not in the repertoire of novice students


and their absence can undermine even the
most sophisticated problem-solving strategies.
Relevant Observations and Experiments
The primary focus of the preceding pages
has been theoretical, concerned with outlining
a prescriptive model of effective human problem solving in a scientific domain such as
physics. In the following paragraphs we mention briefly how this model is related to past
observations and descriptive studies of experts
and novices. Then we point out how such a
prescriptive model can be validated by more
specific experiments.
In some respects the model outlined in the
preceding pages simulates the problem-solving
behavior of expert physicists. For instance,
observations by Larkin and Reif (1979) and
ourselves indicate that experts rapidly
redescribe problems presented to them, often
use qualitative arguments to plan solutions
before elaborating them in greater
mathematical detail, and make many decisions
by first exploring their consequences. Furthermore, the underlying knowledge of such experts appears to be tightly structured in hierarchical fashion.
By contrast, novice students commonly encounter difficulties because they fail to
describe problems adequately. They usually do
little prior planning or qualitative description.
Instead of proceeding by successive
refinements, they try to assemble solutions by
stringing together miscellaneous mathematical
formulas from their repertoire. Furthermore,
their underlying knowledge consists often
largely of a loosely connected collection of such
formulas.
However, as pointed out at the beginning of
this article, our theoretical considerations have
been intended to be prescriptive rather than
descriptive and have thus not primarily aimed
at simulating the problem-solving behavior of
actual experts. Accordingly, we have specified
explicit rules and procedures to accomplish
tasks which experts may actually carry out by
using tacit knowledge or pattern recognition
derived from long experience. For example,
while we have tried to explicate detailed procedures for generating a useful initial description of a problem, experts ordinarily redescribe
a problem in theoretical terms very rapidly and
without conscious effort. While we have

Downloaded by [New York University] at 12:18 23 April 2013

124

F. REIF ND JOAN I. HELLER

formulated explicit application guidelines for


important principles, experts often choose
suitable principles automatically without explicit awareness of the tacit knowledge guiding
their choice.
Observations of experts' behavior are thus
only peripherally relevant to a prescriptive
model of effective human problem solving.
Direct tests of the validity of such a model can
be provided by experiments which deliberately
induce human subjects to act in accordance
with the model and which then observe
whether the resulting performance has the
predicted charateristics and is effective. For example, one can let an individual subject work
problems under carefully controlled conditions
where the subject implements "on-line" external control directions which specify step-bystep how to describe a problem in accordance
with the description procedures outlined in
this article. Indeed, we have recently performed such experiments (Reif & Heller, Note 2;
Heller and Reif, Note 3) which showed that
subjects, guided in this way according to the
model, generate excellent descriptions which
lead subsequently to good problem solutions
(while subjects, working without benefit of the
model, generate descriptions which are often
incomplete or wrong and thus fail to solve
problems correctly.)
Similar experiments can be performed to induce a person to search for problem solutions
in accordance with the methods outlined in
this article. It is also possible to deliberately induce a person to utilize a knowledge base
structured in a particular way. Indeed, Eylon
(1979) and Eylon and Reif (Note 4) have performed such experiments to study forms of
knowledge organization facilitating human
task performance in scientific domains. In all
such experiments, where a person is induced to
act in accordance with a prescriptive model of
good performance, the person's resulting performance can be studied in detail, and protocol data can be collected on the underlying
thought processes verbalized by the person. In
this way one can assess whether various selected
features of the model are sufficient to lead to
good performance. One can also ascertain
whether such features are necessary by comparing the model with alternative models
where these features have been modified or
omitted.
Such experiments, designed to test a prescriptive model of good problem-solving performance, should be carefully distinguished

from experiments designed to test a model of


effective problem-solving instruction. A model
of effective instruction must certainly be based
on a model of good performance which
specifies how the desired ultimate student performance is to be achieved. But it must also incorporate a model describing the student's initial functioning and a model specifying the
instructional transformation process. Furthermore, experiments testing a model of effective
instruction must ensure that students internalize control directions and other knowledge
which can be explicitly externalized in studies
designed to test models of good performance.

Educational Implications
The most common method of teaching
scientific problem solving is to provide
students with sufficient examples and practice.
This method is neither very efficient nor effective. Indeed, as the previously mentioned
observations of novices indicate, most
students' problem-solving skills are quite
primitive and improve only slowly. Furthermore, many students find the problem solving
required in college-level physics courses difficult or even unmanageable. These observations are not too surprising. As our discussion
has shown, the cognitive mechanisms needed
for effective scientific problem solving are
complex and thus not easily learned from mere
examples and practice.
A potentially more effective instructional
method would teach problem-solving skills explicitly on the basis of insights derived from a
model of effective problem solving. Our
discussion in the preceding pages is highly
relevant to such an instructional approach
because it identifies essential components contributing to effective problem solving. In particular, it suggests that each of these separate
components should be taught explicitly and
that all these components should then be integrated to achieve effective problem solving.
For example, our discussion suggests that one
teach separately how to generate good basic
and theoretical descriptions of a problem; how
to analyze a problem qualitatively before its
actual solution; how to search for a problem
solution by decomposing the problem systematically and exploring relevant decisions; and
how to assess the merits of the resulting solution. Similarly, our discussion suggests, in
keeping with suggestions made by others

Downloaded by [New York University] at 12:18 23 April 2013

KNOWLEDGE STRUCTURE AND PROBLEM SOLVING

(Karplus, 1969, 1981), that considerable attention be paid to the organization of the knowledge acquired by students; that students be
taught to structure their knowledge in hierarchical form; and that such knowledge be accompanied by explicit application guidelines.
Our discussion also indicates that some common teaching practices may be dysfunctional
and hinder the development of students'
problem-solving skills. Two examples may suffice: (a) In the quantitative sciences, a desire for
precision often impels instructors and textbooks
to overemphasize mathematical formalism at
the expense of more qualitative modes of
description. As a result, students may even
come to regard such qualitative descriptions as
scientifically illegitimate. Nothing could be further from the truth! As we pointed out, seemingly vague verbal and pictorial descriptions
help greatly in the search and planning of solutions; they are also commonly used by experts.
Hence one needs to teach both qualitative and
quantitative descriptions, and how to use them
jointly in procedures of progressive refinement,
(b) Too little attention is commonly paid to the
organization of the knowledge acquired by
students. For example, knowledge is usually
presented sequentially, locally organized by
topics; but students are given little help to integrate their accumulating knowledge into a
coherent structure facilitating flexible use. Furthermore, scientific arguments or illustrative
problem solutions axe commonly presented in
purely sequential form, rather than in hierarchically structured ways better suited to
facilitate the flexible use of such knowledge
(Eylon & Reif, Note 4).

application guidelines in Table 4 suggest that


the simplest motion principle, Newton's principle mz = F, is the only one of obvious utility
for relating motion and interaction at this
time. The theoretical description of Figure 7
also indicates that all forces on the ball at the
time t2 are directed downward; hence one
chooses to describe the principle mz = F in
terms of its vector components along the
downward direction. Thus one generates the
constraint
ma

2 = Fdown
0-)
where m is the mass of the ball, a2 is the
magnitude of its acceleration at the time t2,
and Fovia is the downward component of the
total force on the ball at this time.
Here the interaction descriptor Fdown is an
unwanted quantity which may be removed by
generating another constraint about it. To
achieve this aim, one uses the application
guideline which suggests that any interaction
descriptor be elaborated by first using the
knowledge about interaction descriptors
(Table 2) and then about interaction laws
(Table 3). In the present case the knowledge of
Table 2 (the "superposition principle") leads
to Fdown = Fg + F s . The knowledge of Table 3
has already been explicated in the theoretical
problem description of Figure 7. It implies
that the gravitational force has a magnitude F g
= mg (where g is the magnitude of the
gravitational acceleration near the earth) and
that the force exerted by the string has a
magnitude F s > 0 (since the string is taut).
Hence Fdown > mg so that the elaborated constraint (1) becomes
<*2>g-

Appendix
Solution of the Pendulum Problem
We outline here the solution of the pendulum problem, previously described in
Figures 6 and 7, to illustrate more fully
the major decision processes involved in its
solution.
The application guidelines in the knowledge
base suggest that the generation of constraints
should ordinarily start with application of a
motion principle. Furthermore, this particular
problem specifies a special condition, the
tautness of the string, at the time t2. Hence
one can usefully begin by applying a motion principle to the ball at this time. The

125

(2)

Here the motion descriptor a2 is again an


unwanted quantity which may be removed by
generating another constraint about it. To
achieve this aim, one uses the application
guideline which suggests that any motion
descriptor be elaborated by using the
knowledge about motion descriptors (Table 1).
In the present case of circular motion, this
table contains the relation ac = v2lr. Applied
to the ball at the time t2, this relation yields
the constraint
a2 = v2V (L.D)

(3)

where v2 is the magnitude of the velocity of the


ball at the time t2.
Here v2 is again an unwanted quantity
which may be removed by generating another

Downloaded by [New York University] at 12:18 23 April 2013

126

F. REIF AND JOAN I. HELLER

constraint about it. None of the information


about the ball's rhotion before the time t2 has
yet been used. In accordance with the
previously discussed application guidelines,
one again tries to generate such a constraint by
applying a motion principle to the ball before
t2. One merely wants to relate the unwanted
speed v2 to other speeds and positions, without
being explicitly interested in any elapsed
time. Thus the output information in Table 4
suggests that the energy principle
A E = WKS is the motion principle likely to be
most useful. An exploration of alternatives,
already discussed as an example in the text,
leads to the decision that this principle is most
usefully applied to the entire time interval between t0 and t2 (rather than between t0 and t\,
and then between t\ and t2).
Following again the application guidelines,
the motion and interaction descriptors in this
energy principle are now systematically
elaborated by using the knowledge in Tables 1
through 3. The knowledge about interaction
descriptors (Table 2) implies that A E = A AT
+ U.U. The knowledge about motion descriptors (Table 1) implies that A / C = Vim v22.
The knowledge of interaction laws (Table 3)
implies that the change in the ball's gravitational potential energy A U = mg (L-2D) and
that the residual work done by the string
T^tcs 0. Thus the elaborated energy principle A = WKS yields the constraint
Vi mv

+ mg(L-2D) = 0.

(4)

The two unwanted quantities a2 and v2 can


now be easily removed by using simple algebra
to combine the contraints (2), (3), and (4). The
result is Z?>(3/5) L. In other words, the answer
to the problem is that the minimum distance
between the hook and the peg must be (3/5) L.
Reference Notes
1. Reif, F. Making scientific concepts and principles effectively usable: Requisite knowledge and teaching implications. Unpublished manuscript, 1982. Available
from F. Reif, University of California, Berkely.
2. Reif, F., & Heller, J. I. Cognitive mechanisms
facilitating human problem solving in physics: Formutation and assessment of a prescriptive model.
Paper presented at the meeting of the American
Educational Research Association, New York, March
1982.
3. Heller, J. I., & Reif, F. Cognitive mechanisms
facilitating human problem solving in physics: Empirical validation of a prescriptive model. Paper
presented at the meeting of the American Educational
Research Association, New York, March 1982.

4. Eylon, B., & Reif, F. Effects of internal knowledge


organization on task performance. Paper presented at
the meeting of the American Educational Research
Association, San Francisco, April 1979.

References
Bhaskar, R., & Simon, H. A. Problem solving in semantically rich domains: An example from engineering
thermodynamics. Cognitive Science, 1977, 1, 193-215.
Brown, J. S., & Burton, R. Multiple representations of
knowledge for tutorial reasoning. In D. G. Bobrow &
A. Collins (Eds.), Representation and understanding,
New York: Academic Press, 1975.
Bundy, A. Will it reach the top? Predictions in the mechanics world. Artificial Intelligence, 1978, 10,
129-146.
Chi, M. T. H., Feltovich, P. J., & Glaser, R. Categorization and representation of physics problems by experts
and novices. Cognitive Science, 1981, 5, 121-152.
Chi, M. T. H., Glaser, R., & Rees, E. Expertise in problem
solving. In R. Sternberg (Ed.), Advances in the
psychology of human intelligence. Hillsdale, N.J.:
Lawrence Erlbaum Associates, 1981.
DcKleer, J. Multiple representations of knowledge in a
mechanics problem solver. Proceedings of the fifth International Joint Conference on Artificial Intelligence.
Cambridge, Mass.: MIT Press, 1977.
Eylon, B. Effects of knowledge organization on task performance. Unpublished doctoral dissertation, University of California, Berkeley, 1979.
Feigenbaum, E. A. The art of artificial intelligence:
Themes and case studies of knowledge engineering.
Proceedings of the Fifth International Conference on
Artificial Intelligence, Cambridge, Mass.: MIT Press,
1977, pp. 1014-1029.
Greeno, J. G. A study of problem solving. In R. Glaser
(Ed.), Advances in instructional psychology (Vol. 1).
Hillsdale, N.J.: Lawrence Erlbaum Associates, 1978.

(a)
Greeno, J. G. Understanding and procedural knowledge
in mathematics instruction. Educational Psychologist,
1978, 12, 262-283. (b)
Hughes, J. K., & Michton.J. I. A structured approach to
programming. Englewood Cliffs, N.J.: Prentice-Hall,
1977.
Karplus, R. Introductory physics: A model approach. New
York: Benjamin, 1969.
Karplus, R. Educational aspects of the structure of physics.
American Journal of Physics, 1981, 49, 238-241.
Larkin, J. H., & Reif, F. Understanding and teaching
problem solving in physics. European Journal of
Science Education, 1979, 1, 191-203.
Larkin, J., McDermott, J., Simon, D. P., & Simon, H. A.
Expert and novice performance in solving physics problems. Science, 1980, 208, 1335-1342.
Newell, A., & Simon, H. A. Human problem solving.
Englewood Cliffs, N.J.: Prentice-Hall, 1972.
Novak, G. Representation of knowledge in a program for
solving physics problems. Proceedings of the Fifth International Joint Conference on Artificial Intelligence.
Cambridge, Mass.: MIT Press, 1977.
Reif, F. Theoretical and educational concerns with problem solving: Bridging the gaps with human cognitive
engineering. In D. T. Tuma and F. Reif (Eds.), Problem solving and education: Issues in teaching and
research. Hillsdale, N.J.: Lawrence Erlbaum
Associates, 1979.

KNOWLEDGE STRUCTURE AND PROBLEM SOLVING

Downloaded by [New York University] at 12:18 23 April 2013

Sacerdoti, E. A structure for plans and behavior. New


York: Elsevier North-Hollard, 1977.
Simon, D. P., & Simon, H. A. Individual differences in
solving physics problems. In R. D. Siegler (Ed.),
Children's thinking: What develops? Hillsdale, N.J.:
Lawrence Erlbaum Associates, 1978.

127

Вам также может понравиться