RD-Ai157 665 HUMAN COGNITION AND PERFORMANCE(U) CALIFORNIA UNIV SAN i/

DIEGO LA JOLLA INST FOR COGNITIVE SCIENCE

NORMAN ET AL. Si MAY 85 N00014-79-C-0323
UNCLASSIFIED F/G 5/10 NL

EhhEEEohEohEEI
EhEEEEEohhhhhI
EEmhEEEEEEohEI
EEEEEmhEEEmhhI
EEEEEEIIEEII
EEEEEEohmhEmhE
 - . . . - . - ',- - ,- . % . - =. % - ' . .= % " . = a . % =

t4;

.
-W3. m2
M m

.
.
ll IL
I..
. 11U

"25
11111
.
J.6
.

"
RESOLUTION TEST CHART
~MICROCOPY
963-A
NATIONALBUREAU OF STANDARDS-
.

A.. =

- . ,,. " . ,-.

,,. . , ,-. .
9.
;-;' . . _ . ..
.- : .- .,';' ,:. .,-,-
0.
=
.. . . . . "; . . ... ;' .-,,.: ",...'
; ;- -; .,. :. ,..,.':
...-. .. '....-.
,' V- .. - ,-',..'.,-..."
.. .: . -.
- ..
.. . . .. . .. . . . .
.. , . .-.. .. - . . ..
CHART.. - :- f.* ".
,- . MICROCOPY RESOLUTION' TEST1963-A~'~ ~is.'dl
I-
. -. dl .G
. . STANDARD-
:..'
",,.''.'
-
,'-.'
i. ;.'-,,' .. :, '. -
.'"'..'
. . . ' . . .
. . - ~ NAIONA OF
BUREAU
.. .....
. .i'
 -v -

A HUMAN COGNITION
LC)
AND PERFORMANCE

ONR Contract N-0014.79-C.0323

NR 667-437 %'.

April 1979-March 196 *S-~'

Final Report

COGNITIVE
SCIENCE

tTlc
ELECTE

INSTITUTE FOR COGNITIVE SCIENCE

UNIVERSITY OF CALIFORNIA, SAN DIEGO LA JOLLA, CALIFORNIA 92093

* d~z4 bu1Ofls
___
Um
__ 7 25 oo
ot.'.

HUMAN COGNITION
AND PERFORMANCE

ONR Contract N-00014-79-C-0323

NR 667-437

April 1979 - March 1985

Final Report

Donald A. Norman and David E. Rumelhart

Principal Investigators

Donald R. Gentner
Skills Project Director

71.

-..'
,...ft'':. Iw, o = ,,h= v ,,,

ml.m documenti h~w been arpprove~d

toe public npm:-- ,rid *~a* IMa
diatrLbution Is unimi di I

%-. ft - "2 ".p\"'

-. - - - - . ., -." ,' -
 Unclassified
SECURITY CLASSIFICATION OF THIS PAGE

REPORT DOCUMENTATION PAGE

la REPORT SECURITY CLASSIFICATION lb. RESTRICTIVE MARKINGS
Unclassified
. 2a. SECURITY CLASSIFICATION AUTHORITY 3. DISTRIBUTION/AVAILABILITY OF REPORT
. /Approved for public release; distribution
2b. DECLASSIFICATIONIDOWNGRADING SCHEDULE unlimited.

4. PERFORMING ORGANIZATION REPORT NUMBER(S) S. MONITORING ORGANIZATION REPORT NUMBER(S)

6a. NAME OF PERFORMING ORGANIZATION 6b. OFFICE SYMBOL 7a. NAME OF MONITORING ORGANIZATION
Institutefor Cognitive Science (If applicable)

6c. ADDRESS (City, State, and ZIPCode) 7b. ADDRESS (City, State, and ZIP Code)
C-015
University of California, San Diego
La Jolla, CA 92093
8a. NAME OF FUNDING /SPONSORING 8ab. OFFICE SYMBOL 9. PROCUREMENT INSTRUMENT IDENTIFICATION NUMBER
ORGANIZATION (If applicable) N-00014-79-C-0323
Research
Personnel & Training
8c. ADDRESS (City, State, and ZIP Code) 10. SOURCE OF FUNDING NUMBERS
Code 442PT PROGRAM PROJECT ]TASK - WORK UNIT
Office of Naval Research ELEMENT NO. NO. NO. ACCESSION NO.
800 North Quincy Street, Arlington, VA 22217 R 667-437
11 TITLE (Include Security Classification)
Human Cognition and Performance: Final Report

12 PERSONAL AUTHOR(S)
Donald A. Norman, David E. Rumelhart, and Donald R. Gentner
13a. TYPE OF REPORT 13b. TIME ,9VRED 14. DATE OF REPORT (Year, Month, Day) 1S. PAGE COUNT
Final Report FROM TO 3/85 1985 May 1 95
16. SUPPLEMENTARY NOTATION

17 COSATI CODES 18. SUBJECT TERMS (Continue on reverse if necessary and identify by block number)
FIELD GROUP SUB-GROUP parallel distributed processing; human-machine interaction
motor skills; learning; interface design; typewriting;
performance
19. ABSTRACT (Continue on reverse if necessary and identify by block number)

This report reviews work performed in the Cognitive Science Laboratory during the
period from April 1, 1979 to March 31, 1985. The work has been focused on three main
areas: parallel distributed processing, human-machine interaction, and skilled human
performance. Summaries of these projects are presented.

20 DISTRIBUTION/ AVAILABILITY OF ABSTRACT 21. ABSTRACT SECURITY CLASSIFICATION

-3UNCLASSIFIED/UNLIMITED 0 SAME AS RPT. 0 DTIC USERS
22a NAME OF RESPONSIBLE INDIVIDUAL 22b. TELEPHONE (include Area Code) 22c. OFFICE SYMBOL

OD FORM 1473,84 MAR 83 APR edition may be used untd exhausted. SECURITY CLASSIFICATION OF THIS PAGE
U CLASSIFId
All other editions are obsolete. Unclassified
.'

D. A. Norman and D. E. Rumelhart i Final Report, NR 667-.437

TABLE OF CONTENTS

INTRODUCTION .................................................................... 1
Overview of the Cognitive Science Laboratory ............................................ 1
The Organization of this Report ...................................................... 2
PARALLEL DISTRIBUTED PROCESSING ................................................ 3
A GENERAL FRAMEWORK FOR PARALLEL DISTRIBUTED PROCESSING .........6
The General Framework ........................................................ 6
A set of processing units. ..........................................
The state of activation. ..................................................... 8
Output of the units. ........................................................ 9
The pattern of connectivity . .................................................. 9
The rule of propagation. .................................................... 11
Activation rule. .......................................................... 11
Modifying patterns of connectivity as a function of experience................... 11
Representation of the envirownent. ............................................. 13
Classes of PDP M odels ............................................................. 13
Paradigmsof Learning ......................................................... 13
HierarchicalOrganizationsof PDP Networks .......................................... 16
Bottom-Up Processing ........................................................... 16
Top-Down Processing ........................................................... 17
Interactive Models ............................................................ 17
Synchronous versus Asynchronous Update ............................................. 18
Specific Versions of the General Parallel Activation M odel .................................. 19
Simple Linear M odels .......................................................... 19
Linear Threshold Units ......................................................... 2D
Brain State in a Box ........................................................... 22
Boltzmann Machines ............................................................ 23
Grossberg ................................................................... 27
Interactive Activation Model ...................................................... 27
Feldman and Ballard ........................................................... 28
Sigma-Pi Units .................................................................. 29
References ...................................................... 31
HUM AN-M ACHINE INTERACTION ...................................................... 32
USER CENTERED SYSTEM DESIGN .................................................... 32
Introduction .................................................................... 34
A Pluralistic Field, a Pluralistic Book .................................................. 34
A Book of Questions, Not of Answers ................................................ 34
Some Possible Dimensions ........................................................ 34
The Structure of the Book .......................................................... 36
Postscript ................................................................... 37
Section I: User Centered System Design ............................................................
Section 11: The Interface Experience .................................................. 39

s"~~~~~~~~.......".
............. ..o. . ."....... .....-... ......... -. . .-......... ..
i.6f

D. A. Norman and D. E. Rumelhart ii Final Report, NR 667437

Section : Users Understanding ....................................................................... .40

Bootstrapping 42
Section IV : User A ctivities ............................................................ 42
Technical Issues .................................................................. 44
- Section V: Toward a Pragmatics of Human-Machine Communication ................... 45
Section V I: Inform ation Flow .......................................................... 46
Help Systems as an Extension of Memory ................................. 48
The Chapters ........................................................................................................ 49
. Section VU : The Contest of Com puting .................................................. 51
References ......................................................................... S3
SKILLED MOTOR PERFORMANCE AT VARIABLE RATES ................................... 55
The Generalized Motor Program ........................................................ 56
Test for a Multiplicative Rate Param eter ................................................. 56
A Simple Mathematical Model ....................................................... 57
Problems With Existing Tests of Multiplicative Rate Parweter.......................... 57
-"- The Constant ProportionTest .......................................... 57
Test Results W ith Sim ulated Data .................................................... 58
Literature Review .................................................................... 59
Laboratory Tasks................................................... 59
Arm movements. ................................................ 59
W rist twist . ................................................................. 62
Key press. .................................................... 62
Locomotion ..................................................................... 63
Cockroach . ........................... ....... .................................................... 63
Lobster . .................................................................... 65
Tortoise. ................................................................... 65
Cat . ....................................................................... 65
D og . ...................................................................... . 65
H uman ..................................................................... 66
Other Highly Practiced Skills ....................................................... 68
Breathing. .................................................................. 68
Speech. ...................................................... 68
Handwriting. ................................................................ 69
Typing . .................................................................... . 70
Music . ..................................................................... 72
Typewriting Data .................................................................... 72
General Method .................................................................. 72
Analyses ........................................................................ 73
Correlationalanalysis. ........................................................ 73
The constant proportiontest. .................................................... 74
Experiments I and 2: Prose ......................................................... 74
Method. .................................................................... 74
Results. .................................................................... . 74
Experiment 3: Random W ords ........................................................ 76
Method. .................................................................... 76

VJ

%aj
 ,- .V

D. A. Norman and D. E. Rumeihart iii Final Report, NR 667437

Results ......................................................................................................... 76
," Experiment 4: Slow, Normal, and Fast Typing 76
Results .. ........................................................................................................
Method ........................................................................................................ 76
-"Patterns of Change W ith Typ ying 77
Rate ....................................................................
77

- S mulation M odel of Typing ................................................................................. . 79

References ...........................................................................................................................
M ultilevel Control of Timing ....................................................................................... 82
l'q
'P U B L I CA TON S ........................................................................................... 84
I...................... . Be
Technical Reports ..................................................... 88
Archival Publications .................................................................................................. 91

Accession For

"'" NTIS CT Ai
DTC T.,'-'

azid/or
jDit t,/1 C1 e

K.

F+.*
 -71 crWr '_2. -W. 7--r

D. A. Norman and D. E. Rumelhart Final Report, NR 667437

INTRODUCTION
This report reviews work performed in the Cognitive Science Laboratoy during the period April 1,
1979-March 31, 1985. The amount of work performed has been considerable and a single overview
would be difficult. Instead, we present reports on the main focii of our efforts. Some work is left
out, not because it was not important and significant, but because it represents more isolated pieces of
research.

Overview of the Cagnitive Science LIaberatary

The Cognitive Science Laboratory is a part of both the nsttue for Cognitive Science and the
Center for Human Information Processing at the University of California, San Diego. Work in the
laboratory is aimed at developing a deeper understanding of the theory and applications of human
information processing. The research aims of the laboratory are broad, but all the work follows a
common theoretical view of cognition, providing a cohesion that binds the separate studies together,
despite the apparent wide differences in content. All of the work can be characterized as being Studies
- of Human Cognition. This is the main focus of the laboratory's efforts, and over the years the work has
covered in-depth problems within the areas of perception, attention, learning, memory, language,
thought, and action. Special emphasis has been placed on knowledge representation, and one major
contribution of the laboratory has been within this area: the representational structure known as Active
Structural Networks. Work in this general area is always in process, elaborated and modulated by the
results and progress of work in the specialty areas. Three specialty areas can be talked about
". separately: parallel distributed processing, skilled human performance, and human-machine interac-
tion"..

Studies of Parallel Distributed Processing Mechanisms of Computation. This work examines a

new conceptualization of processing and representational structures. Under this view, com-
putation takes place through the interaction of large numbers of elementary processing
structures that activate and inhibit one another. The outcome of any computation is the
result of competition and cooperation among a large number of simple processes. The
representation of any specific knowledge is distributed across many structures. Representa-
tions are additive in the sense that different knowledge can be represented in the same set
of processing units, new knowledge being added to the activation values of the old. This
work promises to have major influences on a number of different areas, including basic
models of the human information processing system, issues of representation, and the
development of novel machine architectures for computational processes in general. The
mechanisms can be described as nwaudy4lred, for although they do not directly model
brain structures, their properties am consistent with the neurological evidence. Prtliminary
models have already been applied to our research on typing. reading, information retrieval,
. - and attention.

[V. ..
.7

D. A. Norman and D. E. Rumelhart 2 Final Report, NR 667-437

Studies of Skilled Human Performance. These studies involve the theoretical and cxperimen.
tal investigation of skilled human behavior, including the changes that take place a a per-
son acquires a skill. In recent years the effort has focused upon detailed investigation of
novice and expert typing, but now includes the study of timing, general issues in the learn-
ing and practice of skilled performance, and the role of stress in performance.
* Studies of Humm-Machine Interaction. This area has a its goal the determination of the cog-
nitive principles that underlie human-machine interaction. This work, therefore, merges
traditional work on Human Factors with Cognitive Psychology, the difference being the
emphasis on cognitive and mental processing. As a result, much of the work concentrates
upon the study of mental representations (mental models), memory structures (and limita-
tions), principles of social interaction, and skilled human performance. Some of the work
is based upon our theoretical descriptions of the mechanisms that give rise to human error
and our models of human information processing structures; in addition, there is special
emphasis on the work on parallel distributed systems, the roles of different knowledge
structures, and the competing demands upon the person. The work in this group interacts
heavily with our work in other areas, both by applying the insights gained from the theories
and experimental findings, as well as adding to the other areas, for in the process of
developing a theoretical structure for interaction with machines, it soon becomes obvious
that there are gaps in the general understanding of cognitive functioning, gaps that must be
filled before there can be successful application.

The Organization of this Report

During the contract period both the projects in Parallel Distributed Processing (PDP) and
Human-Machine Interaction (HMI) have completed manuscripts for major books that review the work
of the laboratory and also move the field forward in major ways. The books themselves provide the
best summaries of our efforts in these two areas. Accordingly, we felt the best way of preparing these
sections of the report was to prepare a summary of the books. The review of the PDP effort is
presented by the chapter in the PDP book that best provides the overall summary. It is written by
Rumelhart, Hinton, and McClelland and entitled "A General Framework for Parallel Distributed Pro-
cessing." All three of the authors have contributed to the contract efforts. McClelland was closely
associated with the work at UCSD until he moved to Carnegie-Mellon University at the beginning of
the 1984/85 academic year. Hinton was a postdoctoral fellow in the Cognitive Science program at
UCSD where this work initiated (see the book that resulted from the earlier Cognitive Science pro-
gram at UCSD: Hinton & Anderson, 1981). Hinton is now at Carnegie-Mellon University.

The work on Human-Machine Interaction is summarized through the book introduction by

Drr- and Norman and its section introductions.

The work on study of Skilled Human Performance is summarized in s report prepared by Donald
Oentner.

* . . . . . .,.' =======.======= ====== :. -. " .

 D. A. Norman and D. E. Rumelhart 3 Final Report, NR 667-437

PARALLEL DISTRIBUTED PROCESSING

EXPLORATIONS IN THE MICROSTRUCTURE OF COGNITION
The following is a list of the chapters which will appear in the above titled two volume book to
* be published by NMT Press/Bradford Books, Cambridge, Massachusetts,

VOLUME 1: FOUNDATIONS edited by David E. Rumeihatt and James L. McClellmnd.

1. The PDP Perspective

1. The Appeal of PDP

J. L. McClelland, D. E. Rumelbart, and G. E. Hinton
2. A General Framework for PDP
D. E. Rumelhart, G. E. Hinton, and 1. L. McClelland
3. Distributed Representations
G. E. Hinton, 3. L. MtCltlland, and D). E. Rumihart
4. Discussion & Preview
D. E. Rumeihart and J. L. McClelland

HI. Models of Basic Mechanisms

Iii 5.
Feature Discovery by Competitive Learning
D. E. Rumelhart and D. Zipser
6. Introduction to Harmony Theory (Sinolensky)
P. Smoleusky
7. Learning and Relearning in Boltzmann Machines
I 0. E. Hinton and T. J. Sejnowski
8. Error-Driven Learning in Multilayer Networks
D. E. Rumelhart, G. E. Hinton, and R. 3. Williams
 D. A. Norman and D. E. Rumelbart 4 I'inal Report, NR 667-437

3 Ill. FormalAnalyses o~f PDP Models

9. An Introduction to Lineur Systems

M. 1. Jordan
10. The Logic of Activation Rules
R. J. Williams
11. The Delta Rule
G. Stone
12. P3: A Parallel Network Simulating System
D. Zipser and D. PRabin
13. Resource Requirements of PD? Networks
J. L. McClelland

* VOLUME 11: APPLICATIONS edited by James L. McClelland and David E. Rumelhart

* IV. Models of Psychological Processes

Higher Level Processes

* 14. Schemata
D. E. Rumnelhart, P. Smolensky, J. L. McClelland, G. E. Hinton

Perception, Learning, Memory, and Thought

15. TRACE Model of Speech Perception
J. L. McClelland and J. L. Elman
16. The Programmable Blackboard Model of Reading
J. L. McClelland
17. A Model of Distributed Memory and Amnesia
J. L. McClelland and D. E. Rumeihart
18. On Learning the Past Tenses of English Verbs
D. E. Rumeihart and 3. L. McClelland
19. Thought Processes
D. E. Rumelhart, P. Smolensky, J. L. McClelland, G. E. Hinton
D. A. Norman and D. E. Rumelhart 5 Final Report, NR 667437

V. PDP Models A the Brdn

20. Certain Aspects of the Anatomy and Physiology of the Cerebral Cortex
F. Crick and C. Asanuma
21. Five Questions about Cerebral Cortex
T. J. Sejnowski
22. Neural and Conceptual Interpretation of PDP Models(Smolcnsky)
P. Smolensky
23. Biologically Plausible Models of Place Recognition and Goal Location
D. Zipser
24. Neural Plasticity & Critical Periods
P. W. Munro

Vl. Epilog

25. Reflections: Cognition and Parallel Distributed Processing

D. A. Norman
26. The Future
J. L. McClelland and D. E. Rumelhart

What makes people smarter that computers? This book suggests that the answer lies in the mas-
sively parallel architecture of the human mind. While thought and problem solving have a sequential
character when viewed over a time frame of minutes or hours, we argue that each step in the sequence
is the result of the simultaneous activity of a very large numberr of simple computational elements,
each influencing others and being influenced by them. This book describes our work in developing a
theoretical framework to the development of models of aspects of perception, memory, language, and
thought.

Volume 1 lays the theoretical foundations of parallel distributed processing. Section 1 introduces
the approach and the reasons why we feel it is a fruitful one. Section 2 describes several models of
basic mechanisms with wide applicability to different problems. Section 3 presents a number of
specific technical analyses of different aspects of parallel distributed models.

Volume 2 applies the approach to a number of specific issues in Cognitive Science and Neurosci-
ence. Section 4 describes parallel distributed processing models of thought, speech perception, and
reading, and models of aspects of memory and language acquisition. Section 5 discusses the relation
between parallel distributed processing models and neurophysiology, and describes a number of models
which a specifically addressed to neurophysiological data. Section 6 concludes the work with a discus-
sion of the strengths and weaknesses of the approach and directions for the future.
D. A. Norman and D. E. Rumelhart 6 Final Report, NR 667-437

A GENERAL FRAMEWORK FOR PARALLEL DISTRIBUTED PROCESSING

D. E. Rumelhart, G. E.Hinton, and J. L. McClelland

This chapter provides a general framework for PDP models and then shows properties of certain
specific realizations of the general model.'

The General Franework

It is useful to begin with an analysis of the various components of our models and then describe
the various specific assumptions we can make about these components. Theme are eight major aspects
of a parallel distributed processing model:

(a) a set of processing units

(b) a state of activation
(c) an output function for each unit
(d) a pattern of connectivity among units
(e) a propagation rule for propagating patterns of activities through the network of connectivi-
ties
() an activation rule for combining the inputs impinging on a unit with the current state of
that unit to produce a new level of activation for the unit.
(g) a learning rule whereby patterns of connectivity are modified by experience
(h) an environment within which the system must operate.

Figure 1 illustrates the basic aspects of these systems. There is a set of processing units generally indi-
cated by circles in our diagrams; at each point in time, each unit % has an activation value, denoted in
the diagram as at(t); this activation value is passed through a function f ( to produce an output value
o,(t). This output value can be seen as passing through a set of unidirectional connections (indicated
by lines or arrows in our diagrams) to other units in the system. There is, associated with each connec-
tion, a real number, asually called the weight or strength of the connection designated wj which deter-
mines the amount of effect that the first unit has on the second. All of the inputs must then be com-
bined by some operator (usually addition) and the combined inputs to a unit, along with its current
activation value, determine, via a function F, its new activation value. The figure shows illustrative
examples of the function f and F. Finally, these systems are viewed as being plastic in the sense that
the pattern of interconnections is not fixed for all time. Rather the weights can undergo modification
as a function of experience. In this way the system can evolve. What a unit represents can change
with experience and the system can come to perform in substantially different ways. In the following
sections we develop an explicit notation for each of these components and describe some of the

3. We ae, of couts. not the fint to atteapt a poeral characterization of this pual cleat of abodels. Amah (197) and
Feldman and Ballard (1982) ase two pape witb similarly igecml aims.

.:..:-:.. .. -....... .
: .. :...:- ....... ..-. .... ..-.... .. ............-..-...-.. . ..
D. A. Norman and D. E. Rumelhart 20 Final Report, NR 667-437

The pattern association paradigm is the typical learning situation for a linear model. There is a
set of input units and a set of output units. In general each input unit may be connected to any out-
put unit. Since this is a linear network, there is no feedback in the system nor are there internal units
between the inputs and outputs. There are two sources of input in the system. There are the input
patterns which establish a pattern of activation on the input units and there are the teaching units
which establish a pattern of activation on the output units. Any of several learning rules could be
employed with a linear network such as this, but the most common are the simple Hebbian rule and
the delta rule. The linear model with the simple Hebbian rule is called the simple linear associator (c.f.
Anderson 1970). In this case, the increment in weight w0 is given by Aw1 = 'Iajt. In matrix notation,
this means that AW = rlTAr. The system is then tested by presenting an input pattern without a
teaching input and see how close the pattern generated on the output layer matches the original teach-
ing input. It can be shown that if the input patterns are orthogonal, 4 there will be no interference and
the system will perfectly produce the relevant associated patterns exactly on the output layer. If they
are not orthogonal, however, there will be interference among the input patterns. It is possible to
make a modification in the learning rule and allow a much larger set of possible associations. In par-
ticular, it is possible to build up correct associations among patterns whenever the set of input patterns
are linearly independent. To achieve this an error correcting rule must be employed. The delta rule is
most commonly employed. In this case, the rule becomes Awij = -n(r,-aj)aj. What is learned is essen-
tially the difference between the desired response and that actually attained at unit u due to the input.
Although it may take many presentations of the input pattern set, if the patterns arc linearly indepen-
dent the system will eventually be able to produce the desired outputs.

The examples described above were for the case of the pattern associator. Essentially the same
results hold for the auto-associator version of the linear model. In this case, the input patterns and
the teaching patterns are the same and the input layer and the output layer are also the same. The
tests of the system involve presenting a portion of the input pattern and having the system attempt to
reconstruct the missing parts.

Linear Threshold Units

The weaknesses of purely linear systems can be overcome through the addition of nonlinearities.
Perhaps the simplest of the nonlinear system consists of a network of linear threshold units. The
linear threshold unit is a binary unit whose activation takes on the values {0,1}. The activation value of
unit % is 1 if the weighted sum of its inputs is greater than some threshold 19 and zero otherwise.
The connectivity matrix for a network of such units, like the linear system, is a matrix consisting of
positive and negative numbers. the output function, f, is the identity function so that the output of a
unit is equal to its activation value.

It is useful to see some of the kinds of functions which can be computed with linear threshold
units which cannot be computed with simple linear models. The classic such function is the excluive
or (XOR) illustrated in Figure 4. The idea is to have a system which responds (1) if i teceives a {01)
or a {10} and .sponds (01 otherwise. The figure shows a network capable of this pattern. In this case

4. See Chapter 9 for a diacumion or orthogonaity. linear independence and the like,
 .- - _: _ . -- , -;. .-. .II- II. I - - . :. .. _.. : - .
-.

D. A. Norman and D. E. Rumelhart 19 Final Report, NR 667-437

updating at a time. Among other things, this system can help the stability of the network by keeping
it out of oscillations which are more readily entered into with synchronous update procedures.

Specific Versions of the General Parallel Actlvation Model

In the following sections we will show how specification o, the particular functions involved pro-
duces various of these models.

Simple Linear Models

Perhaps the simplest model of this class is the simple linear model. In the simple linear model
activation values are real numbers without restriction. They can be either positive or negative and are
not bounded. The output function, f (aj) in the linear model is just equal to the activation level aj.
Typically, linear models consist of two sets of units, a set of Inpa units and a set of oiapw units. (As
discussed below, there is no need for internal units since all computation possible with a multiple step
linear system can be done with a single step linear system.) In general, any unit in the input layer may
connect to any unit in the output layer. All connections in a linear model are of the same type. Thus,
only a single connectivity matrix is required. The matrix consists of a set of positive, negative and zero
values, for excitatory, inhibitory values and zero connections, respectively. The new value of activa-
tion of each unit is simply given by the weighted sums of the inputs. For the simple linear model with
connectivity matrix W we have

A (t+1) = WA ().

In general, it can be shown that a linear model such as this has a number of limitations. In par-
ticular, it can be shown that nothing can be computed from two or more steps that cannot be com-
puted by a single step. This follows because the above equation implies

A (t+1)=W 'A (0).

We can see this by proceeding step by step. Clearly,

A (2) = W A (1) = W (WA (0)) = W 2A (0)

It should be clear that similar arguments lead to A (i +1)=W'A (0). From this, it follows that for every
linear model with connectivity matrix W, that can attain a particular state in t steps there is another
linear model, with connectivity matrix W', that can reach the same state in one step. This means,
among other things, that there can never be any computational advantage in a linear model of multiple
step systems, nor can there ever be any advantage for allowing feedback.

. '" - °
 D. A. Norman and D. E. Rumelhart 18 Final Report, NR 667437

Table 3

Leel I units Level 2 units Level 3 units

ul u2 u3 u4 u5 u6 u7 u8 u9 ulO uil u12

Level I ul within levd I

'input' u2 level I affecting
units u3 effects level 2
u4
Leve 2 u5 level 2 within level 2
"intemal' u6 affecting leve 2 affecting
units u7 level I effects level 3
us

Level 3 u9 level 3 within

.output, u1O affecting levd 3
units ull levd 2 effects
u12

It is sometimes supposed that a "single level" system with no hierarchicalstructure in which any
unit can communicate with any other unit is somehow less powerful that these multilevel hierarchical
systems. The present analysis show that, on the contrary, the existence of levels amounts to a restric-
tion on the general of free communication among all units. Such noliderarchical systems actually form
a superset of the kinds of layered systems discussed above. There is, however, something to the view
that having multiple levels can increase the power of certain systems. In particular, a "one-step" system
consisting of only input and output units and no communication between them in which there is no
opportunity for feedback or for "internal" units is less powerful than systems with "internal" units and
with feedback. Since, in general, hierarchical systems involve many "internal" units, some intralevel
communication and some feedback among levels, they are more powerful than systems not involving
such 'internal" units. However, a system with an equal number of "internal" units, but one not charec-
terizable as hierarchical by the communication patterns is, in general, of more potential computational
power. We address the issue of "internal units" and "single step" versus "multiple-step" systerp in our
discussion of specific models below.

Synchronous versus Asynchronous Update

Even given all of the components of the PDP models we hav described so far, there is still
another important issue to be resolved in the development of specific models. That is the timing of
the application of the activation rule. In some models, there is a kind of central timing pulse and after
each such clock tick a new value is determined simultaneously for all units. This is a sychraso ardase
procedure. It is usually viewed a a discrete, difference approximation, to an underlying continuous,
differential equation in which all units are continuously updated. In some models, however, units are
updated asynchronously and at random. The usual assumption is that each at each poino in time each
unit a a fixed probability of evaluating and applying its activation rule and updating its activation
value. This later method ha certain theoretical advantages and was developed by Hopfield (1967) and
has been employed in Chapter 6, Chapter 7, and Chapter 14. The major advantage is that since the
units are independently being updated if we look at a short enough time interval only one unit is

" ...
."" - " .,..' w' "'" " - ' - - '- -- "., .
"-'.. . . - . "- - " - . "- . "- - " ' " " ." "- -'" ""-' " ' " " '" . '"
D. A. Norman and D. E. Rumelhart 17 Final Report, NR 667-437

divided up into 9 regions. The upper-left region represents interactions among level 1 units. The
entries in the upper-middle region of the matrix represents the effects of level 1 units on level 2 units.
The upper-right region represcnts the effects of level 1 units on level 2 units. Often bottom-up models
do not allow units at level i effect units at level i+2. Thus, in the diagram we have left that region
empty representing no effect of level 1 on level 3. It is typical in a bottom-up system to assume as well
that the lowest level units (level 1) are input units and that the highest level units (level 3) are output
units. That is, the lowest level of the system is the only one to receive direct inputs from outside of
this module and only the highest level units affect other units outside of this module.

Top-Down Processing

The generalization to a hierarchical top-down system should be clear enough. Let us order the
units into levels just as before. A top-down model then requires that the upper-right regions of the
weight matrix be empty-that is, no lower level unit affects a higher level unit. Table 2 illustrates a
simple example of a top-down processing system. Note, in this case, we have to assume a top-down

Table 2

Levl I units Levd 2 units Levd 3 units

ul u2 u3 u4 u5 u6 u7lu u9 ul0 ull u2

Level I ul witbin
*input* u2 level I
units u3 effects
u4

Level 2 us level 2 within

'internal' u6 affecting level 2
units u7 level I effects
us

Level 3 u9 levd 3 within

.output, ul0 affecting levd 3
units ull levd 2 effects
u2

input or *message' which is propagated down the system from higher to lower levels as well as any data

input which might be coming directly into level 1 units.

Interactive Models

Interactive models are simply models in which there can be both top down and bottom up con-
nections. Again the generalization is straightforward. In the general interactive model any of the cells
of the weight matrix could be nonzero. The more restricted models in which information flows both
ways, but in which information only flows between adjacent levels assume only that the regions of the
matrix more than one region away from the main diagonal are zero. Table 3 illustrates a simple three
level interactive model with both top-down and bottom-up input. Most of the models that actually
have been suggested count as interactive models in this sense.

----. - - .
D. A. Norman and D. E. Rumelhart 16 Final Report, NR 667437

sometimes called unsupervised learning. Each different kind of unsupervised learning procedure has its
own evaluation function. The particular evaluation procedures are mentioned when we treat these
models. The three unsupervised learning models discussed in this book are addressed in Chapters 5, 6,
and 7.

HierarchicalOrganizationsof PDP Networks

It has become commonplace in cognitive science to describe processes n 'top-down," 'bottom-

up,' interactive" to consist of many stages of processing etc. It is useful to see how these concepts can
be represented in terms of the patterns of connectivity in the PDP framework. It is also useful to get
some feeling for the processing consequences of these various assumptions.

Bottom-Up Processing

The fundamental characteristic of a bottom-up system is that units at level i may not affect the
activity of units at levels lower than i. To see how this maps onto the current formulation it is useful
to partition the coalitions of units into a set of discrete categories corresponding to the levels their
inputs come from. There are assumed to be no coalitions with inputs from more than one level.
Assume that there are L, of units at level i in the system. We then order the units such that those in
level LI are numbered u I ...uL those in level L 2 are numbered uL1+I • • L1+L2 etc. Then, the con-
straint that the system be a pure bottom-up system is equivalent to the constraint that the connectivity
matrix, W, has zero entries for wij in which uj is the member of a level no higher than j. This
amounts to the requirement that the lower left hand region of W contains zero entries. Table 1 shows
this constraint graphicly. The table shows an example of a three level system with four units at each
level. 3 This leads to a 12x12 connectivity matrix and an A vector of length 12. The matrix can be

Table I

Level units Level 2 units Level 3 units

u] u2 u3 u4 u5 u6 u7 ug u9 u]O ul] 12

Level I ul within level I

*input' u2 level I affecting
units u3 effects level 2
u4

Level 2 u5 within level 2

'intema u6 level 2 affecting
units 7 effects level 3
us
Level 3 u9 within
'output' u10 level 3
anits a l] effects
u12

3. In general, of course, we would expect many levels and many units at each level

" '. . . . .
"-."o.. ' - = . . mum
.i ik-a, daa,,i Ivm t, .. s,.I ,.dk kd " , ,.-.'-,. .
 D. A. Norman and D. E. Rumelhart 15 Final Report, NR 667-437

Oupu

Maoli ofofnollte Input Units.eto Output Uniteo

inp.t Mofiiabl
niute& Alweights

*OututUts aeMdfbl
InUnits

Connectivity Matrix
PAto Asocator

fom achinpt raiingpatern ar prsened o bth heinput &n

inptuit.Durng Allpu
weigts h egtcnet
in tO nu t h utput nitsero.Dril
uigti are etpten r rsne ofia
Molfe ntsan

th outut (fterAndeson
uitsisn masurd.
respnse 197) (b Ileconnctivty onnetorvite atrxmcstr h

structure ofaithe
Fiueb.(aodwete pattern bohiptadotu are
eociation
situation. There
ntwo isircshos of un.
Noeihtah eachtoutput unitsond seawl b ooutuf
unitdsetfthoutpu undiits ach nt connectios
is feinputac itews

resposothotptuiei Anderson.itAJns
(after mrd. 977)
1ivrt(d) conectivty mati ofo the att-enoiator.
Thbe utse

- connect to eli other unite with modifiable weights.

i.%
 D. A. Norman and D. E. Rumelhart 14 Final Rteport, NR 667437

(2) Reg darity discovery, in which units learn to respond to 'interesting" patterns in their input.
In general, such a scheme should be able to form the basis for the development of feature
detectors and therefore the basis for knowledge representation in a PDP system.

In certain cases these two modes of learning blend into one another, but it is valuable to see the dif-
ferent goals of the two kinds of learning. Associative learning is employed whenever we are concerned
with storing patterns so that they can be re-evoked in the future. These rules are primarily concerned
with storing the relationships among subpattcrns. Regularity detectors are concerned with the meming
of a single units response. These kinds of rules are used when feare discovery is the emential task at
hand.

The associative learning case generally can be broken down into two subcases-patte associa-
tion and auto-association. A pattern association paradigm is one in which the goal is to build up an
association between patterns defined over one subset of the units and other patterns defined over a
second subset of units. The goal is to find a set of connections so that whenever a particular pattern
reappears on the first set of units, the associated pattern will appear on the second set. In this case,
there is usually a teaching iqvp to the second set of units during training indicating the desired pattern
association. An auo-association paradigm is one in which an input pattern is associated with itself.
The goal here pattern completion. Whenever a portion of the input pattern is presented the remainder
of the pattern is to be filled in or completed. This is similar to simple pattern association, except that
the input pattern plays both the role of the teaching input and of the pattern to be associated. It can
be seen that simple pattern association is a special case of auto-association. Figure 3 illustrates the two
kinds of learning paradigms. Figure 3A shows the basic structure of the pattern association situation.
There are two distinct groups of units-a set of input units and a set of output units. Each input unit
connect with each output units and each output unit receives an input from each input unit. During
training patterns are presented to both the input and output units. The weight connecting the input
to the output units are modified during this period. During test patterns are presented to the input
units and a response on the output units is measured. Figure 3B shows the connectivity matrix for the
pattern associator. The only modifiable connections are from the input units to the output units. All
other connections are fixed at zero. Figure 3C shows the basic structure of the auto-association situa-
tion. All units are both input and output units. The figure shows a group of 6 units feeding back on
itself through modifiable connections. Note that each unit feeds back on itself as well as on each of
its neighbors. Figure 3D shows the connectivity matrix for the auto-associator. All units connect to all
other units with modifiable weights. In the case of auto-association, there is potentially a modifiable
connection from every unit to every other unit. In the case of pattern association, however, the units
are broken into two subpatterns, one representing the input pattern and another representing the
teaching input. The only modifiable connections are those from the input units to the output units
receiving the teaching input. In other cases of associative learning the teaching input be more or less
indirect. The problem of dealing with indirect feedback is difficult, but central to the development of
more sophisticated models of learning. Barto & Sutton (1987) have begun a nice analysis of such learn-
ing situations.

In the case of regularity detectors, a teaching input is not explicitly provided, instead, the teach-
ing function is determined by the unit itself. The form of the internal teaching function and the
nature of its input patterns determine what features the unit will learn to respond to. This is

--: :', ,-.-. --

- '- - ' " .- ;- "".""."
" ,." - - ~~.. . -.
" D. A. Norman and D. E. Rumelhart 13 Final Report, NR 667437

theorem has been proved. Still another variation has

AWgj = 'qa(t)(oj(t)-wqj).

This is a rule employed by Grossberg (1976) and a simple variant of which has been employed by us in
Chapter 5. There are many variations on this generalized rule, and we will describe some of them in
more detail when we discuss various specific models below.

Representation of the environment. It is crucial in the development of any model to have a clear
model of the environment in which the model is to exist. In PDP models, we represent the environ-
ment as a time varying stochastic function over the space of input patterns. That is, we imagine that at
any point in time, the is some probability that any of the possible set of input patterns is impinging on
the input units. This probability function may in general depend on the history of inputs to the sys-
tem as well as outputs of the system. In practice, most PDP models involve a much simpler characteri-
zation of the environment. Typically, the environment is characterized by a stable probability distribu-
tion over the set of possible input patterns independently of past inputs and past responses of the sys-
tem. In this case, we can imagine listing the set of possible inputs to the system and numbering them
from 1 to M. The environment is then characterized by a set of probabilities, pi for I = 1,..,M. Since
each input pattern can be considered a vector it is sometimes useful to characterize those patterns with
nonzero probabilities as constituting "orthogonal" or "linearly independent" sets of vectors, 2 Certain
PDP models are restricted in the kinds of patterns they are able to learn. Some being able to learn to
respond correctly only if the input vectors form an orthogonal set, others if they for a linearly indepen-
dent set of vectors and still others are able to learn to respond to essentially arbitrary patterns of
inputs.

Classes of PDP Modela

There are many paradigms and classes of PDP models that have been developed. In this section
we describe some general classes of assumptions and paradigms. In the following section we describe
some specific PDP models and show their relationships to the general framework outlined here.

Paradigmsof Learning

Although most learning rules have roughly the form indicated above, we can categorize the learn-
ing situation into two distinct sorts. These are:

(1) Associative learning, in which we learn to produce a particular pattern of activation on one
set of units whenever another particular pattern occurs on another set of units. In general,
such a learning scheme must allow an arbitrary pattern on one set of units to produce
another arbitrary pattern on another set of units.

L See Cbeptet 9 tot opication of then term.

o4

%-
 D. A. Norman and D. E. Rumelhart 12 Final Report, NR 667437

(1) The development of new connections.

(2) The loss of existing connections.
(3) The modification of the strengths of connections which already exist.

Very little work has been done on (1) and (2) above. To a first order of approximation, however, (1)
and (2) can be considered a special case of (3). Whenever we change the strength of connection away
from zero to some positive or negative value that has the same effect as growing a new connection.
Whenever we change the strength of a connection to zero that has the same effect u losing an existing
connection. Thus, in this section we will concentrate on rules whereby strengts of connections are
modified through experience.

Virtually all learning rules for models of this type can be considered a variant of the Hesbias
learning rule suggested by Hebb in his classic book Organdzaion of Behavior, (1949). Hebb's basic idea
is this. If a unit, uj, receives a input from another unit, uj, then, if both are highly active, the weight,
w~j, from us to ui should be strengthened. This idea has been extended and modified so that it can be
more generally stated as

Awa, = (a ).t, (t))h(oj (t),W(J),

b,

where ti(t) is a kind of teaching input to *. Simply stated this equation says that the change in the
connection from uj to u, is given by the product of a function, a), of the activation of , and its
teaching input tj and another function, ho, of the output value of Nj and the connection strength
Swi In the simplest versions of Hebbian learning there is no teacher and the function g and h amn sim-
ply proportional to their first arguments. Thus we have

Aw,s =l.atoj,

where il is the constant of proportionality representing the learning rate. Another common variation
is a rule in which h(oj(t),wj) = oj(t) and S(a(t),t(t))= ilQt ()-a(t)). This is often called the
Widrow-Hoff rule. However, we call it the delta rule because the amount of learning is proportional to
the difference (or delta) between the actual activation achieved and the target activation provided by a
teacher. In this case we have

AwIJ = "i (()-a,())o,().

This is a generalization of the pereepron learning rule for which the famous percepim cmergece

it

+-; ~...+. .-........................... .... •......

... ..... ......
D. A. Norman and D. E. Rumelhart 11 Final Report, NR 667437

The pattern of connectivity is very important. It is this pattern which determines what each unit
represents. As we shall see below many of the issues concerning whether 'top-down' or 'bottom-up'
processing systems are correct descriptions or whether a sytem is hierarchical and if so how many lev-
els it has, etc. arm all issues of the nature of the connectivity matrix. One important issue which may
determine both how much information can be stored and how much serial processing the network
must perform, is the fan In and fan ow of a unit. The fm in is the number of elements which either
excite or inhibit a given unit. Thefm ow of a unit is the number of units affected directly by a unit.
Note, in some cases we need more general patterns of connectivity. Specifying such a pattern in the
general case is complex and will be addressed in a later section of this chapter.

The rule of propagation. We also need a rule which takes the output vector, 0(t), reprenting
the output values of the units and combines it with the connectivity matrices to produce a mt ipw for
each type of input into the unit. We let net i be the net input of type i to unit u. Whenever only
one type of connectivity is involved we suppress the first subscript and use netj to mean the net input
into unit uj. In vector notation we can write NET,(t) to represent the net input vector for inputs of
type i. The propagation rule is generally straight forward. For example, if we have two types of con-
nections, inhibitory and excitatory, the net excitatory input is usually simply the weighted sum of the
excitatory inputs to the unit. This is given by the vector product NET, = WO(t). Similarly, the net
inhibitory effect can be written as NET4 = W 1O (). When more complex patterns of connectivity are
involved, more complex rules of propagation are rejuired. We treat this too in the final section of the
chapter.

Activation rule. We also need a rule whereby the net inputs of each type impinging on a particu-
lar unit are combined with one another and with the current state of the unit to produce a new state
of activation. We need a function, F, which takes A(t), and the vectors NETj for each different type
of connection and produce a new state of activation. In the simplest cases, when F is the identity
function and when all connections are of the same types, we can write A(t+l) = WO(t) = NET(t).
Sometimes F is a threshold function so that the net input must exceed some value before contributing
to the new state of activation. Often, the new state of activation depends on the old one as well as
the current input. In general, however, we have

A (t +1) F (A (t).,ET(t),,NET (t)2, " )

the function F itself is what we call the activation rule. Usually, the function is assumed to be deter-
ministic. Thus, for example, if a threshold is involved it may be that a,(t) - I if the total input
exceeds some threshold value and 0 otherwise. Other times it is assumed that F is stochastic. Some-
times activations are assumed to decay slowly with time so that even with no external input the activa-
tion of a unit will simply decay and not go directly to zero. Whenever t%(t) is assumed to take on con-
tinuous values it is common to assume that F is a kind of sigmoid function. In this cue. an individual
unit can irare end reach a minimum or maximum value of activation.

Modifying patterns of connectivity as a fwuction qf experience. Changing the processing or

knowledge structure in a parallel distributed processing model involves modifying the patterns of inter-
connectivity. In principle this can involve three kinds of modifications
 D. A. Norman and D. E. Rumelhart 10 Final Report, NR 667437

Network Representation

\.U U2 U3 U4 u5 uO U7 US

U, 0 0 0 0 +4 0 0 -6

U2
Unpt 0 0 0 0 0 0 -2 +1

uF 0 0 0 0 0 +6 0 0
4 U4 0 0 0 0 +4 -1 +4 0

Us 0 +6 0 -4 0 0 0 0

T Us -5 +5 -3 0 0 0 0 0

U7 -5 0 -2 0 0 0 0 0

9 0+6 +3+ 0 0 0 0

Output Matrix Representation

Figute 2. The connectivity of a network represented by a network drawing and in a matrix. The figre shows an dogt
unit network with units numbered from I to S. Units I to 4 ae input units. They rceive inputs from the outside world mod
feedback from the output unitsc-units 5 through S. The connections among the units ar indicated by the open and filled
disks. The size of the disk indicates the strength of connection. Thus, the large black disk on the line connecting unit I to
unit 8 indicates a strong inhibitory connection from I to S. Similarly, the lSpg open disk on the output line from unit S to
unit 2 indicates that unit 8 strongly excites unit 2. Tle same coonections we shown in the matrix repretatioo on the left.
The +6 in the row for me and the column for U2 indicates that unit S strongly excites unit 2. It will be noted that bheemer
there Is a disk on a line connecting the output of one unit to the input of amother in the network diagam there is a
*.\. corresponding nonzero entry in the matrix. If the disk is filled, the entry in the madi is negatiw. If the disk is open the on-
try is positi. 1he laruger the disk the Pter the .wutude of the entry in the matla. It wigbt also be oted tht the eon-
.* ections In the network ba bean laid out to corrmpond to te entrie of the matd. TMe black disk in the upper left eorn.
of the network corre odB to the -6 in the upper left corer of the matih. Rach disk in the network Is in the cresponding
podtion of its location In the matrix The network would not ham had to be drawn In the way, of cours, and the matriz
would still capture all of the connectivity information in the network. In pnral. becaume network drawings we difficult to
work with we will often imply uae the matrix reprentatio to specify the pattern of connectivity.

- .d *-*
.
D. A. Norman and D. E. Rumelhart 9 Final Report, NR 667437

implications of these various assumptions.

Output of the units. Units interact. They do so by transmitting signals to their neighbors. The
strength of their signals, and therefore the degree to which they affect their neighbors, is determined
by their degree of activation. Associated with each unit, %, there is an output function, f 1(a&()),
which maps the current state of activation a(t) to an output signal o(t) (i.e. o (t) = f ,(a(t))). In
vector notation, we represent the current set of output values by a vector, 0(t). In some of our
models the output level is exactly equal to the activation level of the unit. In this case f is the identity
function f(x) = x. More often, however, f is some sort of threshold function so that a unit has no
affect on another unit unless its activation exceeds a certain value. Sometimes the function t is
assumed to be a stochastic function in which the output of the unit depends in a probabalistic fashion
on its activation values.

The pattern of connectivity. Units are connected to one another. It is this pattern of connec-
tivity which constitutes what the system knows and determines how it will respond to any arbitrary
input. Specifying the processing system and the knowledge encoded therein is, in a parallel distributed
processing model, a matter of specifying this pattern of connectivity among the processing units.

In many cases, we assume that each unit provides an additive contribution to the input of the
units to which it is connected. In such cases, the total input to the unit is simply the weighted sum
of the separate inputs from each of the individual units. That is, the inputs from all of the incoming
units are simply multiplied by a weight and summed to get the overall input to that unit. In this case,
the total pattern of connectivity can be represented by merely specifying the weights for each of the
connections in the system. A positive weight represents an excitatory input and a negative weight
represents an inhibitory input. As mentioned in the previous chapter, it is often convenient to
represent such a pattern of connectivity by a weight matrix W in which the entry w1j represents the
strength and sense of the connection from unit aj to unit %. w Ij is a positive number if unit NJ
excites unit %; it is a negative number if unit uj inhibits unit ut; and it is 0 if unit uj has no direct con-
nect to unit nt . The absolute value of w ij specifies the strength of the conwction.

Figure 2 illustrates the relationship between the connectivity and the weight matrix.

In the general case, however, we require rather more complex patterns of connectivity. A given
unit may receive inputs of different kinds whose affects are separately summated. For example, in the
previous paragraph we assumed that the excitatory and inhibitory connections simply summed algebrai-
cally with positive weights for excitation and negative weights for inhibition. Sometimes, more com-
plex inhibition/excitation combination rules are required. In such caes it is convenient to have
separate connectivity matrices for each kind of connection. Thus, we can represent the pattern of con-
nectivity by a set of connectivity matrices, Wj, one for ach type of connection. It is common, for
example, to have two types of connections in a model, an inhibitory and an zxitatory connection.
When the models assume simple addition of inhibition and excitation they do not constitute different
types in our present sense. They only constitute distinct types when they combine through some more
complex rules.

.. . . ..................................
 D. A. Norman and D. E. Rumelhart 8 Final Report, NR 667437

alternate assumptions which have been made concerning each such component.

A set of processing units. Any parallel activation model begins with a set of processing units.
Specifying the set of processing units and what they represent is typically the first stage of specifying
an PDP model. In some models these units may represent particular conceptual objects such as
features, letters, words or concepts in others thel -r- simply abstract elements over which meaningful
patterns can be defined. When we speak of a distributed representation, we mean one in which the
units represent small, feature-like entities. In this case it is the pattern as a whole which is the mean-
ingful level of aaialysis. T1his should be contrasted to a one unit oe concept representational system in
which single units represent entire concepts or other large meaningful entities.

We let N be the number of units. We can order the units arbitrarily and designate the ith unit
All of the processing of a PDP model is carried out by these units. There is no executive or other
overseer. There are only relatively simple units each doing it own relatively simple job. A unit's job is
simply to receive input from its neighbors and, as a function of the inputs it receives, to compute an
output value which it sends to its neighbors. The system is inherently parallel in that many units can
carry out their computations at the same time.

*-. Within any system we are modeling, it is useful to characterize three types of units Input, ohnsTA
and internal-input units receive inputs from sources external to the system under study. These inputs
may be either sensory input or inputs from other parts of the processing system in which the model is
embedded. The output units send signals out of the system. They may either directly affect motoric sys-
tems or simply influence other systems external to the ones we are modeling. The internal units arc
those whose only inputs and outputs are within the system we are modeling.

The state of activation. In addition, to the set of units, we need a representation of the state of
the system at time t. This is primarily specified by a vector of N real numbers, A(t), representing the
pattern of activation over the set of processing units. Each element of the vector stands for the activa-
tion of one of the units at time t. The activation of unit u, at time t is designated a,(t). It is the pat-
tern of activation over the set of units which captures what the system is representing at any time. It is
useful to see processing in the system as the evolution, through time, of a pattern of activity over the
set of units.

Different models make different assumptions about the activation values a unit is allowed to take
on. Activation values may be continuous or discrete. If they are continuous, they may be unbounded
or bounded. If they are discrete, they may take binary values or any of a small set of values. Thus in
some models units are continuous and may take on any real number as an activation value. In other
eases, they may take on any real value between some minimum and maximum such as, for emple, the
interval (0,I). When activation values are restricted to discrete values they most often ar binary.
Sometimes they are restricted to the values 0 and I where I is usually taken to mean that the unit is
active and a zero is taken to mean that it is inactive. In other models, activation values are restricted
to the values {-I,+I) (often denoted simply I-,+)). Other times nonbinary discrete values are invlved.
7Thus, for example, they may be restricted to the at (-1,0,+1), or to a small finite set of values such as
(1,23,4,5,6,7,8,9). As we shall see, each of these assumptions leads to a model with slightly different
characteristics. It is part of the program of research represented in this book to determine the

[o
*D. A. Norman and D. E. Rumelhart 7 Final Report, NR 667-437

Atli

f (0j)
I /
itg - 1.11j(r)
ntbhold Output
Function Spoid Actimslon
Fuaiction

4,.Fiprc 1. The basi components of a paael distrbuted procoduS systeni

D. A. Norman and D. E. Rumelhart 21 Final Report, NR 667-437

XOR Network

Output
Unit

+1 +1

Internal
Units
Thresholds =.O1

+1 +1

Input Output
Input
Units O00
0

10 1

Figure 4.A network of linear threshold units capable of reponding correctly on the XOR problei.
 D. A. Norman and D. E. Rumelhart 22 Final Report, NR 667-437

we require two layers of units. Each unit has a zero threshold and responds just in case its input is
greater the zero. The weights are ± 1. Since the set of stimulus patterns are not linearly independent
this is a discrimination that can never be made by a simple linear model and cannot be done in a single
step by any network of linear threshold units.

Although multilayered systems of linear threshold units art very powerful, and, in fact, are capa-
ble of computing any boolean function, there is no generally known learning algorithm for this general
case. There is, however, a well understood learning algorithm for the special case of the pereeptron. A
perceptron is essentially a single layer network of linear threshold units without feedback. The learn-
ing situation here is exactly the same as that for the linear model. An input pattern is presented along
with a teaching input. The perceptron learning rule is precisely the same form as the delta rule for
error correcting in the linear model, namely, Awij = i(t -ai)aj. Since the teaching input and the
activation values are only zero or 1, the rule reduces to the statement that

(1) Weights are only changed on a given input line when that line is turned on (i.e. aj = 1).
(2) If the system is correct on unit i (i.e. ii = a1 ) make no change on any of the input weights.
(3) If the unit j responds zero when it should be one, increase weights on all active lines by
amount 7.
(4) If the unit j responds one when it should be zero, decrease weights on all active lines by
amount 71.

There is a theorem, the perceptron convergence theorem, which guarantees that if the set of patterns
are learnable by a perceptron this learning procedure will find a set of weights which allow it to
respond correctly to all input patterns. Unfortunately, even though multilayer linear threshold net-
works are potentially much more powerful than the linear associator, the perceptron for which a learn-
ing result exists can learn no patterns not learnable by the linear associator. It was limitations on what
perceptrons could possibly learn that led to Minsky and Pappert's (1968) pessimistic evaluation of the
perceptron. Unfortunately that evaluation has incorrectly tainted more interesting and powerful net-
works of linear threshold and other nonlinear units. As we shall see in the course of this book the
limitations of the one-step perceptron in no way applies to the more complex networks.

Brain State in a Box

The brain state in a box model was developed by Anderson (1977). This model too is a close rela-
tive of the simple linear associator. There is, however, a maximum and minimum activation value aso-
* ciated with each unit. Typically, units take on activation values in the interval 1-1,1]. The brain state
in a box (BSB) models are organized so that any unit can, in general be connected to any other unit.
The auto-associator illustrated in Figure 3 is the typical learning paradigm for BSB. Note that with this
pattern of interconnections the systems feeds back on itself and thus the activation can recycle through
the system in a positive feedback loop. The positive feedback is especially evident in Anderson &

~~~~~~~~~~~~~~~~~~~~~~~~...
. ......... .. . ..... , .......-...-...-........-. :,,;.,.,..,..,,,--,-,-,.,..
 D. A. Norman and D. E. Rumelhart 23 Final Report, NR 667437

Mozer's (1981) version. Their activation rule is given by

aj(t+1) = aj(t)+,wjjaj(t)

if aj is less than I or greater than -1. Otherwise, if the quantity is greater than 1,aj = 1, and if it is
less than -1, aj = -1. That is,the activation state at time t+1 is given by the sum of the state at time
t and the activation propagated through the connectivity matrix provided that total is in the interval
[-1,1]. Otherwise it simply takes on the maximum or minimum value. This formulation will lead the
system to a state in which all of the units are at either a maximum or minimum value. It is possible to
understand why this is called a brain-state-in-a-box model by considering a geometric representation of
the system. Figure 5 illustrates the "activation space" of a simple BSB system consisting of three units.
Each point in the box corresponds to a particular value of activation on each of the three units. Thus,
each such point corresponds to a state of the system. dimensional space in which the first coordinate
corresponds to the activation value of the first unit, the second coordinate corresponds to the activa-
tion value of the second unit and the third coordinate corresponds to the activation value of the third
unit. Thus, each point in the space corresponds to a possible state of the system. The feature that
each unit is limited to the region [-1,11 means that all points must lie somewhere within the box whose
vertices are given by the points (-1,-1,-l), (-1,-,+-), (-1,+I,-i), (-l,+1,+1), (+1,-1,-l),
(+1,-1,+1), (+,+I,-1), and (+1,+1,+I). Moreover, since the system involves positive feedback, it is
eventually forced to occupy one of these vertices. Thus, the state of the system is constrained to lie
within the box and eventually, as processing continues, is pushed to one of the vertices. Of course, the
same geometric analogy carries over to higher dimensional systems. If there are N units the state of the
system can be characterized as a point within this N dimensional hypercube and eventually the system
ends up in one of the 2 corners of the hypercube.
Learning in the BSB system involves auto-association. In different applications two different
learning rules have been applied. Anderson and Mozer (1981) applied the simplest rule. They simply
allowed the system to settle down and then employed the simple Hebbian learning rule. That is,
AwyJ = -najaj. The error correction rule has also been applied to the BSB model. In this case we use
the input as the teaching input as well as the source of activation to the system. The learning rule thus
becomes Awij = -Q(t -ai)aj where tjis the input to unit i and where a and aj are the activation values
'.. of the system after it has stabilized in one of the corners of the hypercube.

Boltzmann Machines

Another more recent development is the Bolmnam machine developed by Hinton and Sejnowski
(198?). The Boltzmann Machine is discussed fully in Chapter 7, however, it is useful to understand the
basic idea and how it relates to the general class of models under discussion here. To begin, the
Boltzmann machine consists of binary units which take on the values (0,1). The units are divided into
two categories, the visible units corresponding to our input and output units and the idden units
corresponding to our internal units. In general, any unit may connect to any other unit. However,
there is a constraint that the connections must be symmetric. That is, the wij - wjj. In the Boltzmann
Machine, there is no distinction between the output of the unit and its activation value. The activa-

0. " - '" ' '"" ' ". ' ..' .-, . ''-.' ..,. • : "'- , '''.,. "',, : , -"-"-"
 D. A. Norman and D. E. Rumelhart 24 Final Report, NR 667-437

+)

,,A-::: /

AV
- %

E ~/

Figure S. The state space for a three unit version of a BSB modd. E dimnsion of the box repreents the activalon
value of one unit. Each unit is bounded in activation between (-2,11. The curving arrow in the box represnts the meqience of
state the system moved through. It began at the black spot wr the middle of the boz and, a procemin proceeded moved to
.- the (-,+,+) corner of the box. DSB sytems always end up in one or another of the corners. The particular corner depends on
the start state of the network, the input to the system and the patter of connections mon the units.

::,~~~~~~~~~~~~~~~~~..,.
.... _.... ...
.........
.. . . ,...... .,. ....... -.- ,--.... . :...
...
D. A. Norman and D. E. Rumelhart 25 Final Report, NR 667-437

tion values are, however, a stochastic function of the inputs. That is,

p= e -'

where % is the input from outside of system into unit i, 81 is the threshold for the unit and T is a
parameter, called temperature, which determines the slope of the probability function. Figure 6 shows
how the probabilities vary with various values of T. Itshould be noted that as T approaches zero the
individual units become more and more like linear threshold units. In general, if the units exceeds
threshold by a great enough margin it will always attain value 1. If it is fr enough below threshold, it
always takes on value zero. Whenever the unit is above threshold, the probability that it will turn on
is greater than 1/2. Whenever it is below threshold the probability that it will turn off is greater than
1/2. The temperature simply determines the range of uncertainty as to whether it will turn on or off.
This particular configuration of assumptions allows a formal analogy between the Boltzman Machine
and thermodynamics and allows the proof of theorems concerning its performance as a function of the
temperature of the system. This is not the place to discuss these theorems in detail, suffice it to say
that this system, like the BSB system, can be viewed as attaining states on the corners of a hypercube.
There is a global measure of the degree to which each state of the system is consistent with its input.
The system moves into those states which are maximally consistent with the input and with the internal
constraints represented by the weights. It can be shown that as the temperature approaches zero the
probability that the system attains the maximally consistent state approaches 1. These results are dis-
cussed in some detail in Chapter 7.

The learning scheme associated with the Boltzman Machine is somewhat more complex than the
others. In this case the learning events are divided into two classes. During one class a set of patterns
are randomly presented to the visible units and the system is allowed to respond to each in turn. Dur-
ing this phase of learning the system is environmentally driven. During this phase, a simple Hebbian
rule is assumed to apply so that Awj = lata. Note, since activations take on values of zero and 1 this
says that the weight is incremented by an amount -n whenever unit i and j are on, otherwise no change
occurs. During the second phase of learning the system is allowed to respond for an equal period of
time in a so-called free running state in which no inputs are presented. Since the system is stochastic,
it will continue to respond even though no actual stimuli are presented. During this phase, a simple
anti-Hebbian rule is employed, AwJ = -qaiaj. The intuition is roughly that the performance during
the enviromentally driven phase is determined by both the pattern of interconnections and by the
environment. The performance during the free-running phase is determined only by the internal set of
connections. To correctly reflect the environment, we should look at its performance due to the
environment plus internal structure and then subtract out its performance due to internal structure
alone. This is actually quite a powerful learning scheme, it can be shown that if a portion of the input
units are turned on after the system has learned, it will complete the remaining portion of the visible
units with the probability that those units had been present in the stimulus patterns given the subpat-
tern that had been turned on. These issues are again addressed in Chapter 7.
D. A. Norman and D. E. Rumelhart 26 Final Report, NR 667437

1.0
Temperature = 1
o8-

a 0.6

0
&4 0.4-

0.2

0.0 •
-25 -20 -15 -10 -5 0 5 10 15 20 2
Net Input

Figure 6. Probability of sttaining idue I a a function of the distance of the input of the unit from tbhraold. Te
function is ploted for mwal values of T.
* D. A. Norman and D. E. Rumelhart 27 Final Report, NR 667-437

Grossberg

Stephen Grossberg has been one of the major contributors to models of this class over the years.
His work is complex and contains many important details which we cannot review here. We will
instead simply describe some of the central aspects of his work and show how it relates to the general
framework. Perhaps the clearest summary of Crossberg's work appears in Grossberg (190).
Grossberg's units are allowed to take on any real activation value between a minimum a maximum
value. The output function is, in many of Grossberg's applications, a threshold function so that a
given unit will affect another unit only if its activation level is above its threshold. Moreover,
Grossberg argues that the output function must be a sigmoid or S-shaped function of the activation
value of the unit. Grossberg's activation rule is rather more complex than the others we have discussed
thus far in that excititory and inhibitory inputs don't simply sum, but appear separately in the activa-
tion rule. Grossberg has presented a number of possible activation rules, but they typically have the
form

as (t +1) = a,(t)(I-A) + (B -a s (t))net() - (aj(t)+C)neti(t)

where A is the decay rate, B represents the maximal degree of excitation of the unit and C is much
smaller in magnitude than B and represents the maximal amount the unit can be inhibited below the
resting value of 0. Grossberg generally assumes that the inhibitory inputs come from a kind of
recurrent inhibitory field in which the unit is embedded and the excititory inputs come from the unit
itself and from another level of the system.

Grossberg has studied learning in these networks over a number of years and has studied several
different learning schemes. The learning rule he has studied most, however, is similar to the one
analyzed in Chapter 5 and is given by

Awi s = va, (oj -wy).

Grossberg has applied this and similar learning rules in a number of cases, but a review of these appli-
cations is beyond the scope of the present discussion.

Interactive Activation Model

The interactive activation model of McClelland and Rumelhart (1961) and Rumelhart and McClel-
land (192) had units which represented visual features, letters and words. Units could take on any
value in the range Iminmax]. The output function, was a threshold function such that the output was
zero if the activation were below threshold and equal to the difference of the activation value and the
threshold if the activation were above threshold. The interactive activation model involve a connec-
tivity pattern in which units are organized in layers such that an element in a layer connect with exciti-
tory connections with all elements in the layers above and below which am consistent with that unit
and connects negatively to all units in the layers above and below which arm inconnsistent with that
unit. In addition, each unit inhibits all units in its own layer which are inconsistent with the unit in

I-:.
 D. A. Norman and D. E. Rumelhart 28 Final Report, NR 667-437

question. Thus, the interactive activation model is a kind of positive feedback system with maximum
and minimum values for each unit, like the BSB model. The information coming into each units is
weighted (by the interconnection strengths) and summed algebraicly to yield a 'net input' to the unit.
Let net, = 5wjjaj be the net input to unit j. This net input, is then combined with the previous
activation value to produce the new activation value according to the following activation rule.

a+ netj(matr- j )) Meo j>O

aj(t +1)
13 ne:j (aj ( ) -min)) otherwise

where EO is the decay rate of the activation given no input. In words, the new activation value is given
by the old activation value properly decayed, plus (or minus) a factor which pushes toward the
minimum or maximum value depending on the magnitude of the net input into the unit. This activa-
tion rule is similar to that employed by Grossberg, except in this formulation the excitation and inhibi-
tion are algebraically combined.

The interactive activation model was designed as a model for a processing system and our goals
were to show how we could account for specific aspects of word perception. Thus, there was no
specific model of learning proposed to explain where the particular network we assumed came from.
As we shall see, much of the work on learning reported in this book has been aimed at giving plausible
accounts of how such a network might have been learned. (See especially [CompLearn] and [Har-
mony].

Feldman and Ballard

Feldman and Ballard (1982) have proposed a framework for what they dub conrectionism. The
units have continuous activation values, which they call potential which can take on any value in the
range [-1b,10]. Their output function is a kind of threshold function which is allowed to take on a
small number of discrete integer values (OT ol -_9). They have proposed a number of other unit types
each witti a somewhat different activation rule. Their simplest unit type is what they call the P-Uiot.
In this case the activation rule is given by

~+ 1)
a,( =aj (t + Pej(

Once the activation reaches its maximum or minimum value it is simply pinned to that value. Decay is
implemented by self inhibition. Feldman & Ballard also have a covowntive unit similar to our uigma-pi
units described below. Feldman (1981) has also considered learning. In general the approach to learn-
ing offers more machinery than is available within our current framework. In practice, however, the
learning rules actually examined am of the same clam we have already discused.

S . . . i . , . _. , , . . . . . '' . . . '" . . "

'
I - " "" mm :" ' , r a - . '.- * . ,t -- ' -- .
D. A. Norman and D. E. Rumelhart 29 Final Report, NR 667437

Sigma-Pi Units

Before completing our section on a general framework, it should be mentioned that we have
sometimes found it useful to postulate units which are more complex than those described to this
point in this chapter. In our descriptions thus far, we have assumed a simple additive unit in which
the net input to the unit is given by Iwija 1 . This is certainly the most common form in most of our
models. Sometimes, however, we want multiplicative connections in which the output values of two
(or possibly more) units are multiplied before entering into the sum. Such a multiplicative connection
allows one unit to gae another. Thus, if one unit of a multiplicative pair is zero, the other member of
the pair can have no effect, no matter how strong its output. On the other hand, if one unit of a pair
has value I the output of the other is passed unchanged to the receiving unit. Figure 7 illustrates
several such connections. In this case, the input to unit A is the weighted sum of the products of
units B & C and units D & E. The pairs, BC and DE are called coroulcts. In this case we have con-
juncts of size 2. In general, of course, the conjuncts could be of any size. We have no applications,
however, which have required conjuncts larger than size 2. In general, then, we assume that the net
input to a unit is given by the weighted sum of the products of a set of individual inputs. That is the
net input to a unit is given by JwJla,; 1 al ...aj, where i indexes the conjuncts impinging on unit j
and u s, % 2 " " arc the k units in the conjunct. We call units such as these sigma-pi amiss.

In addition to their use as gates, sigma-pi units can be used to convert the output level of a unit
into a signal which acts like a weight connecting two units. Thus, assume we have the pattern of con-
nections illustrated in the figure. Assume further that the weights on those connections are all 1. In
this case, we can use the output levels of units B and D to, in effect, set the weights from C to A and
E to A respectively. Since, in general, it is the weights among the units which determine the behavior
of the network, sigma-pi units allow for a dynamically programable network in which the activation
value of some units determine what another network can do. Indeed, we use this feature to advantage
in Chapter 5.

In addition to its general usefulness in these cases, one might ask whether we might not sometime
need still more complex patterns of interconnections. Interestingly, as described in Chapter 10 we will
never be forced to develop any more complex interconnection type, since sigma-pi units are sufficient
to mimic any function monotonic of its inputs.

t "' '"
"" "''
" " ' '"
".". . "' '" ',i. . ", "'- '.: .--..,-..--.- - . ..
 D. A. Norman and D. E. Rumelhart 30 Final Report, NR 667-437

Sigma Pi Units

ac
IA

to M
D. A. Norman and D. E. Rumelhart 31 Final Report, NR 667-437

References

Anderson, J. A. (19177). Neural models with cognitive implications. In D. LaBerge & S. J. Samuels
(Eds.), Basic processes in reading:Perception and comprehesion. Hillsdale, NJ: Eribaum.

Anderson, J. A., & Mozer, M. (1981). Categorization and selective neurons. In 0. E. Hinton & J. A.
Anderson (Eds.), Parallelmodels o~f associative memory. Hillsdale, Ni: Eribaum.

Anderson, J. A., Silverstein, J. W., Ritz, S. A., & Jones, R. S. (1977). Distinctive features, categori-
cal perception, and probability learning: Some applications of a neural model. Psychsological Review,
84, 413-451.

Barto, Andrew G., & Sutton, Richard (1981). Landmark learning: An illustration of associative
searcb. Biological Cybernetics, 42, 1-8.

Feldman, J. A., & Ballard, D. H. (1982). Connectionist models and their properties. Cognitive Sci-
ence, 6, 205-254.

Grossberg, S. (19176). Adaptive pattern classification and universal recoding, 1: Parallel development
and coding of neural feature detectors. Biological Cybernetics, 23, 121-134.

Grossberg, S. (1980). How does the brain build a cognitive code?. Psychol~ogcal Review, 87, 1-51.

Hebb, D. 0. (1949). The organization o!f behavior. New York: Wiley.

Hinton, G. E., Sejnowski, T. J., & , D. H. Ackley (1984). Boltzmnnu machines: Constraint satisfacuion
networks that learn (Tech. Rep. No. CMU-CS-84- 119). Pittsburgh, PA: Carnegie-Mellon University.

Hopfield, J. J1. (1982). Neural networks and physical systems with emergent collective computational
abilities. Proceedings of the National Academy of Sciences USA, 79, 2554-2558.

McClelland, J. L., & Rumeihart, D. E. (1981). An interactive activation model of context effects in
letter perception: Part 1. An account of basic findings. Psychological Review, 88, 375-4M7.

Minsky, M., & Papert, S. (1969). Perceptrons. Cambridge, MA: MIT Press.

Rumeihart, D. E., & McClelland, J. L. (1982). An interactive activation model of context effects in
letter perception: Part 2. T~he conteitual enhancement effect and some tests and extensions of the
model. Psychological Review, 89, 60-94.
D. A. Norman and D. E. Rumelhart 32 Final Report, NR 667-437

HUMAN-MACHINE INTERACTION

USER CENTERED SYSTEM DESIGN

NEW PERSPECTIVES ON HUMAN-COMPUTER INTERACTION
The following is a list of chapters to appear in the above titled book edited by Donald A. Nor-
man and Stephen W. Draper to be published by Lawrence Erlbaum Associates, Hillsdale, New Jersey.

Introduction
Stephen W. Draper and Donald A. Norman
(With a postscript by Clayton Lewis)

I User Centered System Design

I Architectural Design: An Analogy

Kristina Hooper
2 Issues in Design: Some Notes
Liam J. Bannon
3 Cognitive Engineering
Donald A. Norman

1 The Interface Experience

4 Interface as Mimesis
Brenda K. Laurel
5 Direct Manipulation Interfaces
Edwin Hutchins. James D. Hollan. and Donald A. Norman
6 Notes on the Future of Programming
Andrea A. diSessa

I Users' Understandings

7 User Understanding
Mary Riley
8 Understanding What's Happening in System Interactions
Clayton Lewis
9 Naive Theories of Computation
David Owen
10 Models of Computation
Andrea A. diSessa
11 Knowledge-Based Interface Design
William Mark

• ............
....................-
,. ',..:q........,.-.: .;. ': .,',.'....:........,-...,,.,................-..,-......".......,...-......-.-..

"-::-"
" "-d'-"
"t" "' -&...... ..........
u"- ' ''~mnh.,.n-...-,,a." ''"
D. A. Norman and D. E. Rumclhart 33 Final Report, NR 667-437

IV User Activities

12 The Structure of Users' Activities

Allen Cypher
13 Psychological Issues in Support of Multiple Activities
Yoshiro Nfiyata and Donald A. Norman
14 Communication Paradigms for Window Systems
Rachel Rrichnian (Adar)

V Toward a Pragmatics of Humnan-Machine Commualcatias

15 There's More to Interaction Than Meets the Eye: Some Issues in

Manual Input
William Buxton
16 Display Managers as the Basis for User-Machine Communication
Stephen W. Draper

VI Information Flow
17 Answers First, Then Questions
David Owen
18 Hclping Users Help Themselves
Claire O'Malley
19 Helping Users Help One Another
Liam J. Bannon
20 Designing for Error
Clayton Lewis and Donald A. Norman
21 Computer-Mediated Communication
Liam J. Bannon

VII The Context of Computing

22 From Cognitive to Social Ergonomics and Beyond

John Seely Brown
D. A. Norman and D. E. Rumelhart 34 Final Report, NR 667-437

Introduction

A Pluralistic Field, a Pluralistic Book

What is this field of Human-Computer Interaction? People ae quite different from computers.
This is hardly a novel observation, but whenever people use computers, there is necessarily a zone of
mutual accommodation and this defines our area of interest. People arm so adaptable that they are
capable of shouldering the entire burden of accommodation to an artifact, but skillful designers make
large parts of this burden vanish by adapting the artifact to its users. To understand successful design
requires an understanding of the technology, the person, and their mutual interaction, and that is what
this book is about.

The computer can be thought of from the perspective of its technology--from the field of com-
puter science. Or it can be thought of as a social tool, a structure that will change social interaction
and social policy, for better or for worse. It can be thought of as a personal assistant, where the goals
and intentions of the user become of primary concern. It can be viewed from the experience of the
user, a view that changes considerably with the task, the person, the design of the system. The field of
human-computer interaction needs all these views, all these issues, and more besides. Studying these
various perspectives can involve many disciplines: computer science, psychology, artificial intelligence,
linguistics, anthropology, and sociology-the cognitive sciences. We are prepared to take on board any
discipline, any approach that helps. It is a pluralistic field, so this is a pluralistic book.

A Book of Questions, Not of Answers

This is a book about the design of computers, but from the user's point of view: User Centered
System Design. The emphasis is on people, rather than technology, although the powers and limits of
contemporary machines are considered in order to know how to take that next step from today's lim-
ited machines toward more user-centered ones. This book is about the directions in which we must
move to get there and does not follow traditional paradigms from any of the contributing disciplines.
It is not a book on how to do things. It does not cover techniques and tools. Nor is it a book of fan-
tasies, of possible dream worlds. Instead, it treads an intermediate path, neither detailed science nor
flights of fancy. Think of it as a book of ideas, of analytical techniques described for their purpose,
not for the details of their methods. Think of it as a book from which to derive the new directions
in which we must move. Think of this as a book of questions, not a book of answers.

Some Possible Dimensions

This book is primarily an expression of a pluralistic approach, but if it has a common theme--a
unity in its diversity-it is that human-computer interface design is not one small aspect of the main
business of software design, nor will it be illuminated (let alone 'solved') by a single methodology or
technical innovation. To begin with, we do not wish to ask how to improve upon an interface to a
program whose function and even implementation has already been decided. We wish to attempt User
Centered System Design, to ask what the goals and needs of the users are, what tools they need, what
kind of tasks they wish to perform, and what methods they would prefer to use. We would like to
D. A. Norman and D. E. Rumelhart 48 Final Report, NR 667-437

The primary value of adopting an information flow perspective is that it causes us to step baL.
from any prior assumptions about the pre-eminent importance of learning, of conceptual knowledge,
or of the role of certain 'official" channels such as manuals and tutorials. Instead the perspective
encourages us to look at the amount of information vital to users that is picked up, used, and forgot-
ten as well as at the importance of sources such as colleagues that do so much of the work supposedly
done by manuals. Hence, the information flow perspective suggests that a designer's task in this area is
first to calculate what information must be conveyed to users (the information content needed) and
then to orchestrate the possible information sources to achieve this, perhaps designing special delivery
systems. Some people promote slogans or codes of practice such as "good systems must have on-line
help" and "always write a manual entry." Others have proposed rules for screen design, technical writ-
ing styles, and so on. Before we are ready to consider these assumptions and approaches to the effec-
tiveness of particular information delivery systems, we should look at the broader perspective of the
information flows actually important in current systems--both their content and the sources now sup-
plying them.

A second aspect of the information flow perspective is to consider opportunities for replacing a
requirement for users to know something by timely delivery of the needed information. In the sim-
plest case, a menu might list available commands. However, there is also scope for aiding not just sim-
ple knowledge of items like names, but also understanding. By providing examples ('recipes") an
action can be described to the user in such a way that it can be performed even if the user understands
little about it. This is important whenever completion of an interaction is more important than educa-
tion of the user. Thus, timely information delivery can stand in for memory, saving planning and
understanding. This view can be developed by regarding the design of help systems as an attempt to
design an extension of the user's memory.

Help Systems as an Extension of Memory

A simple but interesting viewpoint for considering help systems is to regard them as memory aids,
for conserving the user's memory resources. Thus, we could consider all help systems as forming an
extended external memory for the user-one kind of mind amplifier. From the cognitive point of view
there is a spectrum of retrieval types. The best situation is when users can remember the item needed
effortlessly. The next best is when they find the needed item already displayed. Third best is when
they must ask. It is only third best because this counts as an interrupting activity, causing suspension
of their current activity (see the discussions in Section IV of this book). This is the distinction made
in the literature between active and passive help-between whether the user has to prompt delivery or
not. Fourth best (i.e., worst) is when the user doesn't know how to ask. The activity requiring the
information will have to be suspended indefinitely, and either a random search will ensue or the user
must just wait to stumble on the answer. This is analogous to when you know you should know
something--a name perhaps-but you just can't recall it. The information is there, but inaccessible.

The ideal, in fact, is not just to reduce the effort of seeking information but (subjectively) to
abolish it-to achieve where possible the smooth flow of information. A familiar instance of this
occurs in every display editor which revolves around the manipulation of the cursor's position. Users
seldom note the subtask of retrieving the cursor position. Instead, they simply edit; moving the cursor
to wherever it should go. Adequate information flows from the display at every moment. We should
D. A. Norman and D. E. Rumelhart 47 Final Report, NR 667-437

what they know by an active use of external sources of information (Draper 1984). Thus information
flow remains important for all users, although a shift is to be expected in the relative importance of
different sources of information as a user gains experience: from a dependence on other people to a
facility for interrogating manuals and source code, and for learning from experimentation.

There are two main topics for study here: identifying the information content and identifying
information sources or delivery systems.

Examples of kinds of Information content:

File names,
Little recipes for the key strokes to get some neat or useful effect,
Getting into a system or subsystem (e.g., how to invoke an editor)
A minimum set of commands
Advanced commands
Where to get information
Efficiency--neater ways to do certain things
How things work
Strategies for combining system programs (functions) for user tasks
Recognizing information when it's there
How to use information you have
How to recover from errors
What is commonly done on this system (i.e., standard user tasks and standard plans for accom-
plishing them)

Eamples of Information delivery systems:

Printed manual entries

On-line manual entries
"Cheat Sheets" (summaries of commands)
Menus
Owen's "DYK" System ('Did You Know': See Chapter 17)
People-replying to specific queries
People-fortuitous pickup (looking over someone else's shoulder)
Monitors that examine your input and periodically make suggestions about new commands, etc.

There are an unlimited number of possible delivery systems, since they can vary in who supplies the
information (a public bulletin board versus system designer documentation), in whether the informa-
tion is specifically requested or spontaneously generated (error mesags, monitors), in the sczess struc-
tures (i.e., by command name, English description, etc.), in how much information about the user is
consulted in selecting the presentation (none, or records of past requests, of commands vocabulary
displayed in practice, etc.).

' ' ' " ''

I " " '" " - " " " - '" ' i " " -i . : ' " : " ' " ', " " '. . ". .' .L -. .t.- . .- .. ' .. :_ .
 D. A. Norman and D. E. Rumelhart 46 Final Report, NR 667437

In Chapter 16, Draper begins by introducing the term "inter-referential [/0 ° as a generalization of
the ways in which the inputs and outputs to a system can refer to each other. This allows for a com-
mon language and history between a user and a system, allowing both to refer to the same concepts
readily and easily. These are critical aspects of the 'first-personness" concept of Laurel (Chapter 4) and
the Direct Engagement notion of Hutchins, Hollan, and Norman (Chapter 5). Draper suggests that
systematic support for inter-referential I/O could have important benefits. First, it could provide
better tools for support of user-machine dialogue. Second, it should allow for error correction and
recovery through cooperative interaction, by allowing the user to elicit clarifications and explanations.
These are points taken up by other chapters: The ability to interact gracefully with the system to take
care of error is an important component of the discussion by Lewis and Norman in Chapter 20 and in
the discussion of *repair" by Brown in Chapter 22.

Draper's chapter, like that by Buxton, again points out the interaction between apparently low-
level details and high-levcl concepts. On one hand, the concept of inter-referential 1O is a low-level
suggestion (keep the records needed to interpret pointing to objects on the screen and other references
to previous 110) but, on the other hand, it adds up to a transformation of the whole quality of the
interaction and supports repair to the dialogue and smooth transitions between "levels."

Thesc chapters can be seen as steps towards a pragmatics of human machine interaction. The
notion however cries out for further exploration: There may be a whole perspective (not just hints)
waiting to emerge on the "pragmatics" or "ethnomethodology" of human-computer communication that
would explain what are now isolated, curious observations.

Section VI: Information Flow

Information Flow is meant to be a suggestive term, representing the flow of information that takes
place among the systems, documents, and users of a system. At the least, the term "information flow"
is used to emphasize the importance of a unified approach that embraces information sources other
than manuals, including other people, system displays, all system memages, and prompts. In line with
the emphasis of this book on new perspectives, the chapters in this section emphasize relatively
neglected aspects of information flow.

Many issues in the field of Human-Computer Interaction revolve around the flow of information
within the system comprising user(s), machine, and the relevant surroundings (which generally include
printed manuals, phone lines to other users or experts, and colleagues near or far). On a fine scale
information flow analysis concentrates on all the items of information that must reach users during the
execution of their given goals in order for them to be achieved. On a broader scale, information flow
analysis concentrates on all the flows of information that affect a user over time.

It is obvious that information flow analysis includes for all users attention to feedback, such as
the echoing of characters and the movement of cursors, and, for novices, how they pick up all the
necessary facts. With large and complex systems, even very experienced users know only a minority of
the large set of commands: Their expertise lies, not in their having learned enough to immediately
solve any problem, but in their having become skilled in gathering information and supplementing

-°
. .-
. . .""" " "' '2"
D. A. Norman and D. E. Rumelhart 45 Final Report, NR 667-437

Section V: Toward a Pragmatics of Human-Machlne Communication

In linguistics it is traditional to separate the analysis of language into semantics and syntax.
Those who study language, however, have long noted interesting phenomena that do not fall neatly
under these divisions (and which are less amenable to formal analyses). These additional subject areas
are known by various names, but the term " pragmatics' seems most common. Pragmatics addresses
issues such as why the response 'Yes is not appropriate as an answer to the question 'Can you pass the
salt?" and why the response 'I'm not wearing a watch" can be appropriate as an answer to the question
'What is the time?' Analogous issues arise in human-computer communication, and so we suggest
labeling them with the same term: pragmatics.I

The two chapters in this section, in effect, deal with pragmatics, although neither mentions the
term. Other chapters too are connected, most notably the previous chapter, Reichman's "Communica-
tion Paradigms for Window Systems,' with its concern for how a person might signal shifts between
activities smoothly and naturally. A characteristic of pragmatic phenomena is that, although they
often are concerned with 'large scale' structures such as user activities or conversation, they are typi-
cally expressed in "low-level* things such as mouse clicks, or words such as 'but" and 'hello.' Don't be
fooled: "But" and 'hello' may seem non-entities syntactically and have insignificant meaning, but their
role in structuring a conversation is crucial.

In Chapter 15, Buxton demonstrates that critical attention to the details of the input mechanisms
are essential if we are to optimize interface design. He points out that current systems are designed
with a paucity of input devices--often only a keyboard-thereby severely restricting their effectiveness.
He goes on to discuss how even apparently small variations within a class of input device changes their
effectiveness for different tasks considerably. The chapter might at first appear to be about the
mechanics of input devices, but it deals with more than that. The chapter is really about the pragmat-
ics, about subtle features that can seriously affect the nature of the interaction.

Buxton demonstrates quite convincingly that too little attention has been placed on the varieties
of input devices (existing and potential) and on the critical details of their operation. He draws upon
his considerable expertise as a professional musician (especially electronic music) to argue that we
should realize that people have five major output mechanisms which could therefore use five input
devices: two hands, two feet, and voice. There is no reason why we could not use all five simultane-
ously, and he suggests that at least we should develop systems that encourage use of the two hands
simultaneously, each on a different form of device. (We could perhaps add a sizth--eye movements--
but effective eye movement sensors are not yet practical.) Buxton restricts himself to simple kinds of
inputs. He does not discuss the important variety of output devices that are possible, and the paucity
of experimentation with sound as output (not speech-sound) is much to be regretted (see, for ezam-
ple, the PhD thesis by Bly (1982). Although Buxton is fully aware of the power of audio signals as
data, he deliberately restricted the topics in this chapter to make his points: (a) We have seriously
neglected the details and pragmatics of input devices; (b) Details do matter; and (c) We can do better,
much better.

1. te term "pra atics" bM been proposed before in this connection (see Bmtoo, 1991) but the idea seem not to bave yt
pited a foothold.

2' . . .....---- -- ,-" -. - . .

 D. A. Norman and D. E. Rumelhart 44 Final Report, NR 667437

to current window managers, would support users' new conceptualizations of the human-machine
interface.

Technical Issues

All three chapters depend on the fact that many systems today make a start toward supporting
multiple user activities. The most frequently discussed technique is the use of windows. However win-
dows are not, technically, the heart of the matter.

One important issue is whether one user may have more than one interactive process active at
once. Multiple computer jobs are clearly possible, for that is the essence of time-sharing, but with
most of today's smaller, personal computers, only one interactive task can be run at a time. Multiple
processing is possible on systems such as Berkeley UNIX that permit 'suspending," "backgrounding," or
"detaching" a process and then "resuming," "foregrounding,' or "reattaching' to it later, thus allowing
one user at a single terminal to alternate among several interactive processes without terminating and
re-starting them. The great advantage is that the system context of each process is preserved. Windows
add the preservation of the current output state of each process, which can then serve to preserve some
of the user's mental context for the associated activity. Windows also serve as a visual reminder of the
existence and identity of unfinished activities.

In the world of personal computers, windows face stiff competition from "integrated" systems,
systems that achieve by other means the most important aims of a window system-to make it easy, for
users whose complex tasks require multiple sub-activities, to switch rapidly and easily among multiple
parts (programs) of the system, transferring the results of one system into the data structures of
another. Whether the system has windows or not, the overall activity will only succeed if the two pro-
grams share a common representation with which to pass data back and forth-integrated" systems
stress this inter-program data transfer compatibility. Although windows are a leading technique in sup-
porting multiple user activities, they are neither necessary nor sufficient, and, as discussed in
Reichman's chapter, they are not yet supporting users' activities correctly.

From the computer science point of view, the two technical issues a designer must be concerned
with in order to support multiple concurrent user activities on one system are, one, allowing the user
to move between interactive programs without losing the context built up within each (this is espe-
cially important when the two programs are not part of the same user activity), and, second, ensuring
that separate programs can be used together to pursue a single activity with an easy transfer of data
among them. From the psychological point of view, the technical issues concern support for interrup-
tions, reminders, and resumption of activities, all of which could be viewed as support for the natural
and efficient saving and resumption of context. The chapters in this section provide valuable insights
into the nature of real activities and into what will be required to support them in a smooth, even if
tacit, manner.

- "' - * 4 "
i .=. . ' " " ".- "-.. . . . . . . -. . . .
''' . . . . . . . ." "' '
D. A. Norman and D. E. Rumelhart 43 Final Report, NR 667-437

To our knowledge, these aspects of real tasks have seldom been discussed, let alone supported
(or, in the case of displacement activities, discouraged) by computer systems. The three chapters in
this section discuss these aspects of real user activities and suggest several different ways in which com-
puter systems might support the performance of multiple activities. Cypher discusses the situation
from the computer's point of view, showing how system support could be provided. Miyata and Nor-
man discuss the issues from the psychological point of view. And Reichman shows how the analogy to
similar situations in conversation provides useful insight into the problem and shows how analyses of
conversational gambits can suggest techniques for allowing people to switch among tasks and, perhaps
more important, to resume where things left off with a minumum of disruption.

The three chapters in this section are motivated by the understanding that people usually have
many things to do, some of which are related to each other and some which are not. Some tasks take
such a long time that other things are also done along the way. Unexpected events occur frequently,
interrupting on-going activities. As a result, normal activities are filled with simultaneous or over-
lapped tasks, with things deferred until later, or planned tasks awaiting an appropriate time to be
done. These factors make behavior complex, even if each activity is relatively simple.

The basic phenomena of interleaved tasks and the basic modes of computer support for people's
activity structures are introduced in Chapter 12 by Cypher, The Structwe of Users' Activities. Cypher
approaches the problem from the point of view of the user who is faced with the need to deal with
multiple activities that simultaneously compete for attention. He treats the competing activities as
interruptions of the ongoing activity and discusses ways to aid the user to transform the simultaneous
demands into a linear sequence that can be dealt with one at a time. He discusses the varying ways in
which activities may need to share contexts, and establishes in detail the point that people's computer
usage is not, in general, in a simple one-to-one relationship with their activities. Good support for
activities must therefore allow flexible groupings of processes. He then makes some proposals for
activity support in interfaces, particularly in connection with the function of reminding.

The chapter by Miyata and Norman, Chapter 13, supplements the one by Cypher, expanding upon
some of the issues and pointing out the relevance to some traditional studies within psychology.
Cypher's analysis was based upon the activity flow to both the user and the computer. Miyata and
Norman take the psychological point of view, centering upon what is going on inside the person. They
focus more heavily on humans' abilities to pursue multiple activities and the effects of the relevant cog-
nitive limitations. They thus develop a more detailed analysis of reminding, including the distinction
between the signaling component and the descriptive component of a reminder, and hence arrive at
proposals for supporting this in interfaces.

Reichman, in Chapter 14, analyzes how users use window systems in an attempt to support their
multiple activities. She discusses the developmental trend of computer interfaces leading up to win-
dows that has influenced users to expect more from their systems and to act as if they were having a
two-party conversation with the computer. In conversation we switch topics, usually signaling this to
the other person by simple clue words such as well anyway," without recourse to explicit meta-
linguistic references to topic. Reichman argues that interfaces need to achieve a parallel smoothness of
use and that the lack of such mechanisms in existing systems shows up in various annoying and per-
sistent errors. She then presents a set of contextual, navigational, and visual techniques that, if added

' " &i .i -"- - -' -.

". . - '. ". .- " . .' '----- " i "
D. A. Norman and D. E. Rumelhart 42 Final Report, NR 667-437

explore consequences of choices of concepts. It can furthermore support the interpretation of user
inputs.

Bootstrapping

Many of the chapters in this section touch on the concept of oostrapping when dealing with how
new users can get started at acquiring understanding. The phrase 'pulling yourself up by your own
bootstraps" applies with some force to the problem facing new users. It is of course physically impossi-
ble to lift oneself by one's own bootstraps, and there is a corresponding philosophical problem of how
one can come to know anything without first already knowing something. However, many of the
problems that new users face hinge upon the fact that in order to understand the new concepts they
have been exposed to, they need already to understand part of those concepts. All the chapters in this
section are concerned with the kinds of understanding that users must and do acquire, but it is the
consideration of new users that forces the issue of "bootstrapping.' For instance, Owen and Lewis, in
different ways, are both concerned with the way users interpret their experience of events (and "errors')
at the interface. On the one hand, experience can lead to new understanding and a change in con-
cepts. On the other hand, the interpretation of events and, hence, the possibility of new understand-
ing, depends in part on the user's existing concepts. Owen suggests that the first seed of a new piece
of understanding must be delivered "Answers first, then questions": The initial information must be
delivered as an unsolicited display, because the user doesn't yet know enough to formulate and pursue
precise questions.

Section IV: User Activities

The point of the three chapters in this section is that people interleave their activities--and do so
in many different ways. Real tasks can take hours, weeks, or even months to perform. As a result,
other tasks must of necessity intervene. Some tasks are so important they are allowed to interrupt oth-
ers. Some tasks require the generation of subtasks, sometimes to satisfy prerequisite conditions, some-
times as corequisites, sometimes simply as relevant, but not necessary components. And then, some
tasks are so onerous that people, sometimes deliberately, allow displacement activities to take over.

On a broad scale, people and computers often do several things at once: A person can cook
dinner and carry on a conversation; A computer can execute multiple processes (time-sharing) while
performing input-output operations on its peripherals. However both computers and humans use an
important part of their capacities linearizing-carrying on several tasks by switching among them, so that
at any instant only one of them is actually being done, even though on a broader scale they are all
being done in parallel.

Linearizing carries with it a bookkeeping burden. Any activity carries with it a context'--all the
things that are particular to this instance of the activity. Each time an activity is suspended this con-
tet must be stored: Each time it is resumed, the context must be recalled. Thus, linearizing requires
support for remembering and recalling. In addition, the fact that an activity is still waiting for comple-
tion must itself be remembered and recalled.

. . .,.
D. A. Norman and D. E. Rumelhart 41 Final Report, NR 667-437

concerns how well the representation reflects the actual behavior of the system, and integration reflects
to what extent is the representation is tied to other components of a user's knowledge. Riley is con-
cerned with the nature of the understanding that people form of the systems they are learning. A
major issue is how to present the information so that it becomes understood-which is where the con-
cepts of internal coherence, validity, and integration become important. Her other concern is with the
robustness of the knowledge so formed, robust in the sense that it will not collapse in the face of
unexpected outcomes.

The discussion by Lewis (Chapter 8) draws attention to the role of explanations and their con-
struction in understanding and learning about a system. Lewis emphasizes false explanations, explana-
tions that satisfy the learners that they truly understand and follow what is happening, but that in fact
sometimes bear suprisingly little relationship to the truth. In some sense, Lewis shows what happens
when the knowledge structures are too robust-so robust that they survive and prosper despite repeated
evidence that they are erroneous.

Owen (Chapter 9) proposes that we should consider the idea of a 'Nalve Theory of Computa-
tion," analogous to the "naive physics" that underlies our understanding of physical events. This is
another way of characterizing a kind of informal knowledge underlying a user's practical understand-
ing. He also stresses that the problem of how to help users acquire it may be by making things visible:
by "Answers First then Questions" (a suggestion he follows up in detail in Chapter 17) and by Direct
Manipulation or Visible Computation design techniques that make a system self-revealing, rather than
relying on users taking the initiative and asking for explanations.

diSessa (Chapter 10) offers a typology of the kinds of understanding users may have--structural,"
"functional," and "distributed." He stresses that understanding is, in practice, often much less sys-
tematic than may be presupposed by researchers, and the examples he gives of each model type illus-
trate this. diSessa also shows that the common notion in Human-Computer Interaction of a "mental
model" is much too simple: There are different kinds of models, each kind having quite different func-
tional and computational properties.

Several chapters in the book talk about the need for a good, consistent design model and the
need for it to be reflected in a consistent, coherent System Image (see especially Chapters 3 by Norman
and 7 by Riley). Mark (Chapter 11) accepts this challenge, but points out that this consistency is diffi-
cult to achieve unaided, especially when the system to be designed is large and complex, involving a
team of programmers working over a period measured in months or Years. As a result, he proposes a
design tool, one based upon modern knowledge representation techniques from Artificial Intelligence
(the knowledge representation language KL-ONE by Brachman, 1978).

Mark starts frowr the assumption that a consistent system image is a good thing. He describes a
system that helps programmers develop an explicit and consistent "system image model'-a formal
model of the behavior of a system as observable from the user interface. The subsequent implementa-
tion is then constrained by requiring programmers to link each concept in the model to a specific
software data object: The relationship between model concepts must hold between the software
objects. The knowledge representation technique is used to define similarity of concepts as a basis for
consistency in the interface, and its implementation gives a support system to allow designers to

did"
. D. A. Norman and D. E. Rumelhart 40 Final Report, NR 667437

We somehow have to separate the concepts of immediate visual feedback from other components
of interactive computing. Visual programming and immediate representation are of crucial importance,
but this is not all there is to the experience of directness. The nature of the interaction between one's
thoughts and the structures available in the computing environment are also crucial. Hutchins, Hollan,
and Norman speak of the Gulfs of Execution and Evaluation, gulfs comprised of the distances between
mental concepts and the physical requirements and displays of the computational system. diSessa
shows how different programming environments offer quite different tools and powers to the user.
Surface appearance is only a small part of the story, although it is an essential component.

The phrase "direct manipulation" has become a catchword, connoting great ease of use, visual
depiction of the events upon the screen, and specification of computer operation by moving icons
around on a screen, perhaps supplemented by pop-up menus. But both chapters in the section that
analyze the concept-Chapter 5 by Hutchins, Hollan, and Norman and Chapter 6 by diSessa--show
this view to be simplistic. Hutchins, Hollan, and Norman point out that the concept of directness is
complex, involving different concepts for the directness of specifying the action to be performed and
the process of evaluating the result. Moreover, there is a directness between a person's intended mean-
ing and the commands and displays available on the system ("semantic directness') and a directness of
the actions that are to to be performed and the input actions required or the depiction of the result
on the displays ('articulatory directness!). Finally, the feeling of "direct engagement," the concept of
which Laurel speaks and which, implicitly, is what is believed by many to be the hallmark of "direct
manipulation," occurs only under special conditions, when both the Gulfs of Execution and Evalua-
tion are bridged, and when there is inter-referential constancy between the objects manipulated at the
input and viewed at the output, i.e., input and output representations are unified.

The nature of programming environments is further explored in the chapter by diSessa, who
examines in detail the underlying principles of programming, from conventional languages to "device
programming" and "direct manipulation." diSessa illustrates the way that programming languages bor-
row from and count upon analogies with natural language and experience, illustrating his point with
reference to the language, Boxer, he has helped develop at MIT.

Section M: Users' Understandings

This section concerns the understanding users can have of a system. It starts with three chapters
* that attempt to characterize the nature of understanding relevant to a user of an interface: Chapters 7,
* 8, and 9 by Riley, Lewis, and Owen. Then we turn to diSessa and his focus upon the (non-
computational) sources of knowledge brought to bear by users. He discusses three types of under-
- "standing: structural, functional, and distributed. These three types of understanding differ in their
inherent nature and in the applications to which they mainly relate. We conclude the action with a
chapter that applies ideas on the nature of understanding to interface design: the chapter by Mark.
This is only one dimension underlying the topic however. Others include the origin of users's under-
standing, its nature, and how to support it in system design.

Riley (Chapter 7) argues that knowledge relevant to the understanding of a topic has three major
.- components: internal coherence, validity, and integration. Internal coherence reflects how well the
components of the representation that is to be learned are related in an integrated structure, validity

[o-- ~j.
 D. A. Norman and D. E. Rumelhart 39 Final Report, NR 667437

are precise enough to lead to design rules. The trick is to develop the engineering guidelines in such a
way so as not to lose sight of the overall goals of cooperative, enjoyable interaction-of pleasurable
engagement, to use the term coined by Laurel in Chapter 5. Norman's chapter introduces a wide range
of topics that are dealt with further in subsequent chapters.

Section U: The Interface Experience

This section of the book contains chapters that get directly at the question of the quality of the
user's experience. This is of course the ultimate criterion of User Centered System Design, but most
workers approach it obliquely in various ways such as exploring the implementation techniques, or
applying existing cognitive concepts. These chapters attempt more direct analyses. The spirit of the
enterprise is expressed by the image of computer usage described by Laurel in her chapter: Iuerface as
Mimesis. Laurel suggests we think of the computer as a stage, capable of letting the user experience a
world. What we want is to figure out how to design things so that we, as users, can enter the world.
When we read or watch a play or movie, we do not think of ourselves as interpreting light images, we
become a part of the action: we imagine ourselves in the scenes being depicted. We have a "first per-
son experience." So too with well constructed video arcade games and with simulation tools. Well,
why not with computers?

In this section we start with Laurel's notion of Mimesis, a concept borrowed from that under-
credited developer of computational concepts, Aristotle. Aristotle was a dreamer, but like proper
dreamers, he provided deep analyses of his ideas in ways that are fruitful to consider today, a mere M3O0
years later. Hutchins, Hollan, and Norman follow up with an analysis of a new form of interface, one
that Ben Shneiderman has called "direct manipulation," an interface idea made practical by the folks at
the Xerox Palo Alto Research Center, and made popular by the folks at Apple, in the Macintosh com-
puter. Direct manipulation offers great promise of being a route toward Laurel's "first person experi-
ence," but it also provides traps for the unwary. The Hutchins, Hollan, and Norman chapter offers
detailed analyses of what might really be going on. After that diSessa, speaking from his experience in
the design of languages for children of all ages--first LOGO, then Boxer-discusses the future of pro-
gramming, and the directions in which the programming world might move.

There have been major changes in interaction in computers in general and with programming in
particular, with the advent of highly interactive, personal computers. One change came about naturally
with the introduction of screen editors (text editors that provide automatic and full display of the con-
tents of the text being manipulated). These editors instantly reveal the changes to the text, thus giving
the illusion of working directly with the objects on the screen, an illusion that is more difficult to sus-
tain with earlier generations of line or character-oriented editors. A similar revolution is occurring in
the development of computer systems and in programming, perhaps most prominently with the pro-
gramming conventions used in spreadsheets. Again, the hallmark is continual visual presence, a self-
revealing display showing at all times the state of the system. Not very many programming systems have
all these properties. Important examples are the Smalltalk programming system (Goldberg, 1984) and
Boxer, the system described by diSessa.

,12

n "T~ - . - .. - ..-
 D. A. Norman and D. E. Rumelhart 38 Final Report, NR 667-437

As the title indicates, this book is a collection of a variety of new ideas about human-computer
interaction: new interpretations of phenomena, ideas about phenomena in other areas that may be
.- ;-- -relevant here, and (yes) some ideas about widgets, or at least widgets-in-the-large, such as knowledge-
representation technology. We don't think much of this will be of direct use in design, unlike wisdom

about methods, rules, and widgets. Rather, we hope it will indicate new directions that design might
take, or that experimental designs might explore.

We think this is important. Empirical methods and quantitative rules don't expand the design
space in which we work (except when empirical study reveals a problem that we really ae forced to
solve). Riding the flow of widgets takes us into new territory, but not necessarily the territory we
want to be in. We hope some of the ideas we discuss here will lead to new kinds of interfaces, kinds
that are new in some way that we want.

These ideas are not intended to supplant the familiar wisdom. With some differences in balance,
we all endorse all of those viewpoints on design, and the forms of knowledge that go with them. No
one should design an interface without knowing the technology, without thinking about keystroke
-"counts, or without planning for measurement and redesign. So this book is a supplement, not a sub-
stitute.

ayton Lewis

Section I: User Centered System Design

Chapters 1 and 2 of this section raise some issues concerning the nature of design. The essay by
Hooper (Chapter 1) draws a number of interesting parallels between the design of architectural and
computational artifacts and searches for some lessons we might leam from the successes and, perhaps
more important, the failures of the older design profession. The chapter does not provide a list of prin-
*ciples of design or even a set of guidelines for design that can be adopted by members of the human-
computer interaction community, but it does show that 'movements! and slogans in design can become
overarching, with the result that the ultimate purpose of the design artifact is overlooked. Concern
for the needs of the users should be primary.

The short note by Bannon (Chapter 2) expands on this issue, arguing for a socially conscious
design process that is embedded in a full understanding of the users' needs. He looks at the interae-
tion between the design and use of technology, stressing that artifacts can and do have an effect on the
people who use or inhabit them and that building usable and useful systems requires that attention be
paid to user needs at a variety of levels, from ergonomic concerns to organizational issues. The use of
design as a social and cultural force is a contentious issue raised in both these chapters and returned to
in the concluding chapter by Brown.

In the last chapter of this section, Norman discusses the applied side of Cognitive Science, what
he calls Cognitive Engineering. He is concerned with putting together the parts of the formal science
that will make a science of interface design possible. This requires, says Norman, some formal models
of people and of interaction, models that need only be approximations of the ultimate theory, but that

. • ".-..-..--.° .......
."-'-. -.-.. ,- ..-....-....-. -."- ..-. ..-...-.-.-...-.
".".-.-"..---. '..-.-..'-. ..
. D. A. Norman and D. E. Rumelhart 37 Final Report, NR 667-437

*and the interactions between users (some of the chapters on information flow) are considered. Finally,
we finish by considering computer tools as they fit into and affect the social structures within which
users live and work-the last chapter of Section VI by Bannon and the concluding chapter of the
book, the chapter by Brown in Section VII.

Stephen W. Draper
Donald A. Norman

Postscript

A user interface functions at the intersection of many different kinds of things: people,
machines, tasks, groups of people, groups of machines, and more. These different things contribute
diverse constraints and opportunities to the design process. None of them, not even machines, is cap-
tured by adequate theories, that is, theories adequate to tell the designer what's going to happen when
a design is put to work. Two results of this state of affairs are that design is very hard, and that there
are many ideas about how it should be approached.

One view, argued by Gould and Lewis (1985), is that design must be treated as fundamentally
empirical. Designers must work hard to learn as much as possible about the users of the system and
the work they will do with it. They must assume that their initial design ideas, even given this back-
ground information, will be wrong, and plan for repeated redesign. They must base these redesigns on
empirical measurements of the success of the design, made on actual use of an implementation, a pro-
totype, or mockup.

A second view, represented in the influential work of Card, Moran, and Newell (1963). is that the
chaos of interface design can be reduced over time to a kind of approximate order, in which quantita-
tive rules are discovered that make useful predictions about the good and bad features of potential
design.

A third view, perhaps never stated but implicit in most real design work, is that interface design
is driven by technology. Improvements in interfaces stem from improvements in the size and quality of
displays, or the invention of new widgets for moving cursors, or new ways to arrange information on
the screen. Designers should keep up with the news and think about how to ezploit these new possi-
bilities.

Each of these viewpoints has its corresponding design wisdom. Empirical design needs methods:
ways to learn about users, ways to detect, measure, and dianose the problems in a design. Card,
Moran, and Newell want those quantitative rules. Widgeteers want to know about new widgets. But
this book supplies very little of any of this.

-1
,, . ' . . . . , - . . . - .- . - , . . , . . - , - . . - . -., ., .. ', , -. . : . . -. - .- . . ."., . . . ..
 D. A. Norman ar D. E. Rumelhart 36 Final Report, NR 667-437

intentions and thoughts. On a larger scale are the chapters on users' activity structures and on the
nature of interaction among people, between people, between people and systems, and systems and
S- society.

- These different kinds of analysis are often associated with yet another cause for multiple
approaches: differences in methodology, for example in the use of experimental techniques, of imple-
menting trial systems, observing and questioning users, or simply doing theoretical analyses. For
instance, Mark's chapter revolves around an implementation, whereas that by Miyata and Norman
- involves applying existing work in psychology to the concerns of Human-Computer Interaction. But
the chapters do not stress methodology, and we rejected the idea of organizing by methodological
"-. dimensions.

The Structure of the Book

Following the chapters in Section I that survey the field as a whole, we jump to what we see as
medium scale approaches. This is Section II, 'he Interface Experience.' The studies in this section
are traditional in that they focus on one user, one task, and one machine, but their emphasis on the
subjective experience of the user is not conventional. This section deals with a phenomenologically
oriented approach that directly asks the ultimately central question 'what is the experience like for the
user?' In the end, that is the basic question underlying all user-centered design. However, asking it
does not necessarily or directly lead either to useful theory or to practical design methods which is why
- - only a few chapters take this approach.
* The next section of the book, Section III on 'Users' Understandings,' we move to a finer scale.
- One approach in this section is to start with a design or implementation technique and pursue its
consequences which, of course, one hopes will ultimately have beneficial effects for the user. For
instance, in his chapter, Mark describes an aid for designers that ensures a consistent use of concepts in
an interface. That this will benefit users is assumed rather than directly investigated in that work. Yet
another kind of approach in this section offers various ideas about the nature and content of how a
-user comes to understand a system. The hope, of course, is that better theories will lead to mor
informed and hence better interface design, but again that is an indirect consequence. This section of
the book in fact shows still further diversity in the kinds of argument found to be interesting:
Whereas some authors (e.g., Riley) concentrate on the kinds of content users' understandings can have,
diSessa concentrates on the different origins that understanding can have (e.g., loose analogies with
natural language, as opposed to a detailed 'functional' model from a different instantiation of the
same task domain).

We next move to the larger scale chapters, going from an analysis of user activity structures (Sec-
tion IV), to the pragmatic aspects of human-machine interaction (Section V), to the considerations of
the flow of information among people and between person and systems (Section VI), and then to the
societal and cultural impact of design (Section VII). This approach has let us zoom progressively out-
ward to larger and larger scales in which multiple tasks are considered (the section on user activities)

0i ' i: ii i i - - L .i i / .- ii, i. ;.i. i . : -? ' - :i :' ,--: -

 D. A. Norman and D. E. Rumelhart 35 Final Report, NR 667-437

start with the users, and to v-nrk from there.

Granted the premise of User Centered System Design, though, what follows? The more we study
it, the bigger the subject seems to become. Pluralism is the result of the piecemeal recognition of more
and more important aspects to the subject. We are at the point (in the mid 1980s) of realizing just
. how much bigger the problem is than has usually been acknowledged, but we are not within sight of a
grand synthesis or a unifying theory. This book offers "perspectives'--pluralistic voices laying claim to
your attention. The authors contributing to this book interacted to a considerable extent during its
writing. As a result, many mutual connections have been found and are mentioned in the chapters,
but nothing like a single synthesis has yet been constructed. The main message remains that pluralism
is necessary and appropriate at this stage of the field. The chapters reflect this pluralism implicitly, not
by design.

This book has been difficult to organize, for the essence of pluralism is that there is no single
best approach, no single dominant dimension, no single approach that fully organizes the issues to be
faced or the solutions and perspectives that are offered.

One approach, a rather traditional one, is to start with considerations of the person, the study of
the Human Information Processing Structures, and from this to develop the appropriate dimensions of
the user interface. Many of the chapters in this book can be considered to be developments of this
approach.

Another approach is to examine the subjective experience of the user and how it might be
enhanced. When we read or watch a play or movie, we do not think of ourselves as interpreting light
images, we become a part of the action: We imagine ourselves in the scenes being depicted. We have a
"first person experience." So too with well-constructed video arcade games and with simulation tools.
Well, why not with computers? The ideal associated with this approach is the feeling of "direct
engagement," the feeling that the computer is invisible, not even there, but rather what is present is the
world we are exploring, be that world music, art, words, business, mathematics, literature-whatever
your imagination and task provide you.

A different approach would be to focus upon the social context of computing, on the fact that
computers are meant to be tools for carrying out tasks, tasks that are normally done in association
with (or for the benefit of) other people. This approach focuses upon social interaction, upon the
nature of the workplace, about the kinds of assistance people get from one another rather than from
manuals or formal instruction, and on the ways that we might aid and assist this process. This is a per-
spective in which the issue is the use that can be made of interfaces in designing human work, jobs,
offices, social interactions-even society itself-i.e., a complete subordination of interface design to
social concerns.

All of the above approaches are included in this book. After considering a half dozen ways of
organizing the chapters we decided to group the chapters in terms of the scale or grain of analysis
adopted. We see approaches that focus on one user pursuing one task on one machine, a representing
only one point on a range of possible scales of analysis. On a finer scale we have studies that examine
*-• hand motions, keystrokes, and the information-processing structures of actions and perceptions,
 D. A. Norman and D. E. Rumelhart 49 Final Report, NR 667437

aim for a similar subjective effect in delivering all information. Currently not enough is understood to
tackle this systematically.

The cursor example depends on the display being constantly updated without explicit commands
from the user-the technique of aiaodisplay. Autodisplay poses two major problems: what to display,
and whether the user will pick up the information displayed. The compiler cample illustrates the first
problem. Since error codes are useful only as pointers to other better descriptions of the problem, it
is a safe bet to print the latter. On the other hand automatically displaying a piece of source code or a
possible diagnosis is only as good as the reliability of the compiler at knowing which line of the code
needs to be changed or knowing that a set of symptoms has only one possible cause. The second prob-
lem may be called "receptivity"-whether the user will pick up relevant information if it is displayed.
One of us once watched the other discover that half his files seemed to be missing. It took several
minutes before someone else pointed out that he was logged on to a different machine than he
thought. Yet all the time the name of the machine was being displayed in every prompt, only three
character widths to the left of the fixation point as he frantically typed commands. The machine name
was not noticed, even though the user had himself designed that prompt with exactly this sort of prob-
lem in mind. This is a common kind of problem: failing to see a solution that is literally staring us in
the face, even when we are actively searching for a solution. When the information presented does not
fit the current hypothesis it may be ignored. Until we understand what governs receptivity a lot better,
designing information flow will be a patchy business at best, and it will be hard to design information
delivery systems that make users feel that the system's help facilities are an extension of their own
memories.

The Chapters

Owen, in Chapter 17, reminds us that we often learn by first discovering information, then deter-
mining what question it answers-the inverse of the what is thought to be the standard method of
seeking information. Owen calls this approach "answers first, then questions." The point is that on
many occasions we come across an idea or piece of information more or less by chance, and then recog-
nize it as interesting or relevant in some way: i.e., an answer to a dormant question. He illustrates this
mode of information flow in everyday life and presents a fantasy of one style of advanced graphics-
intensive user interface. He describes the exploration of this method in an experimental facility he
calls DYK ("Did you Know?') DYK offers assorted factual tidbits on demand, providing a novel, but
quite effective, method of conveying some kinds of information to the user population. (Furthermore
his DYK facility is fun to use, making it an attractive displacement activity as well as a valuable source
of information about computer usage.)

Another reason for the importance of "answers first, then questions" as an information delivery
mode emerges from considering a tacit assumption behind most conventional help systems. The most
common way to store and present information follows a database model: all the stored items have the
same structure but different details. An example is the telephone directory: All entries (items) show a
name, an address, and a phone number. Computer databases follow this rule, with standardized for-
mats for its records. Computer help systems and manuals also usually have this form, with one entry
per command and various standard parts to each entry.

V. .- . • . ..

) ., ' , ' - °%
. , •. - . . . , . . . .. . .
 D. A. Norman and D. E. Rumelhart 50 Final Report, NR 667437

To ask a question of such systems-to formulate a query-enquirers must know how to

transform their original need for information into queries the system can respond to. This requires a
knowledge of the query language syntax, of the structure of the database, and of the indexing rystem.
For instance, to find a phone number you must know the person's name: in the case of standard tele-
phone directories, knowing only the address will not work because they are indexed by name. A first
improvement of retrieval systems is to provide more and different indices to the same database-for
instance, computer system manuals can to be indexed by function as well as by command name. These
improvements, however, do not overcome the pr'oblem that the enquirer must know the standard
structure of the information and the legal queries, in order to use the index or phone directory. (We
should remember also that many people prefer to ask another rather than to use a directory
themselves--phone companies mount active campaigns *o move their customers away from asking
information operators for help.)

SO'Malley (Chapter 18) concentrates on the more usual case when the user initiates a query to the
system: "questions first, then answers." She asks why users so often prefer to ask other people for help
rathcr than to use manuals. She suggests that a large part of the answer stems from the fact that to
formulate a question for such systems, you already need to know many things. 2 This means that asking
a question is in principle a multi-step process. Failure in any step means failure of the whole. The
apparent superiority of human help is often due to its robustness-the people one consults can aid in
filling in the missing steps and in reformulating the question in an appropriate fashion. Thus in both
measuring the success of a delivery system, and in designing improved ones, we need to identify all the
- steps and ensure that they are all supported by one means or another. It is not enough to provide
answers, we must also support the formulation of questions.

Bannon focuses upon a relatively neglected aspect of information flow in human-machine

interaction-the role of social interactions. In his first chapter in this section, Chapter 19, Bannon
focuses on user-user information flows, arguing that they will always be an important part of the total
pattern of information flow within the larger system of computers plus user community, despite
present and future advances in system-to-user 'ktlivery methods. He suggests that facilities could and
should be installed on computers to facilitate (not to replace) user-to-user flows-on-line human help,
as it were. This chapter is supplementary to the one by O'Malley, for it argues that human assistance is
such a pervasive aspect of information finding and computer usage that systems ought to be designed
with this in mind, to supplement and to extend the ability to get help from others.

4- Lewis and Norman (Chapter 20) discuss the possible responses of a system when it detects a prob-
lem: a class of system-to-user information flow, initiated by the system. This chapter is a product of
the intensive workshop we held to review draft chapters of the book. There, Lewis and Norman
discovered their common interests in the manner that contemporary systems handle cases where the sys-
tem cannot interpret the information given to it. "Error,* says the system: "naughty user." In fact, the
error is just as much on the part of the system for failure to understand as it is on the part of the user.
6 The main point is that there is a breakdown in the pattern of information flow. In normal conversa-
tion, as Lewis and Norman point out listener and speaker do not jump to assign blame for a failure to

2. As MIyake and Norman (1979) once put it in the title of their paper: "To Ask a Ouestion. One Must Know Enough to
Know What Is Not Known.'

'o. "

6-
. .. . . - - .. .... . --- . ..-. .... ) .... .. . -.- .'. . -" - .. .. . ..-- ..- .,. -,... . - . .. -. . .
D. A. Norman and D. E. Rumelhart 51 Final Report, NR 667-437

understand. Instead, the listener simply does the best job possible at forming an appropriate interpre-
tation and seeks help when that is not possible. Lewis and Norman discuss the range of responses
open to the system and its designers to errors by the user. (These are points raised again in Chapter 22
by Brown).

In the last chapter of this section (Chapter 21), Bannon again focuses on user-user information
flows, but in a different sense than in his earlier chapter: not as part of information flow abose the Sys-
tern, supplementing the documentation, but as an end in itself, for which the computer merely acts as
a medium. He reviews some of the most common forms of computer-mediated communication (e.g.,
electronic mail and computer conferencing), and then some ways of distinguishing the effects the
media have on the human interactions they support. He also shows how one might design an office by
deliberately selecting some set of communication tools to fulfill the needs of the office members. This
chapter adopts the largest scale of any in this section: that of augmenting and shaping human interac-
tions and environments by the selection of technology. Bannon reminds us of the social component
and consequences of human-computer interaction and discusses ways that the computer can be used as
a tool to improve and enhance social interaction. All too often today the computer is thought of as
an individual tool, used by a person in relative isolation from all others. But just the reverse is really
the case. First, we need other people in order to learn how to use computers most effectively (the
point of almost all the chapters in this section). Second, computers provide enhanced communication
abilities, both synchronous and asynchronous. And third, the computer is a new tool that can
enhance participatory work, allowing for shared workspaces, and shared communication channels, so
that our ideas can immediately be made available to the others with whom we work, allowing for
increased interaction and creativity.

Section VII: The Context of Computing

This final section has only one chapter, one that in part functions as a sort of summary state-
ment. We organized the book by "grain size," by the level at which the topics were approached. We
think it only fitting that we conclude the book at the largest grain size, with a chapter that considers
the way one might use computational tools to affect society, to change the tasks that we do and the
way that we view them. This section, and the single chapter within it, address many of the issues
raised in the book, but with the explicit attempt to exploit them by designing tools for society, tools
that could change thinking, problem solving, writing, and communication. A good way to put this
chapter into perspective is by reviewing the entire book, commenting upon how Brown uses the points
developed in the book in his far-ranging analysis of social settings and in his suggestions for social
tools.

Section I provides an overview of User Centered System Design. Hooper (Chapter 1) reviews the
design philosophy of architecture, suggesting that we in Human.Computer Interaction can learn from
the history of that field. Bannon (Chapter 2) raises severral issues concerning the role of the designer
in society and the need for a responsible design. In Chapter 22, Brown continues this discussion. He
suggests that the Bauhaus School in architecture, the influential design school in Germany just before
the second world war that took social change as an essential theme of design, might provide a good
model for our field as well. The Bauhaus school and its impact was raised and discussed by both
Hooper and Bannon, but not necessarily with great favor. Brown asks 'how designers who take up this

-!, .
D. A. Norman and D. E. Rumelhart 52 Final Report, NR 667-437

challenge might form a movement akin to the Bauhaus School of architecture, in which physical
artifacts were designed explicitly to facilitate social change." This is both an exciting challenge and a
major problem. But as the experience with the Pruitt Igoe Housing Project in St. Louis vividly points
out (see Chapters I and 2), it is dangerous to make social policy statements prematurely. We certainly
do not know enough about social factors to understand the full implications of the artifacts that we
create. Bannon summarized the issue this way (in an electronic mail interchange among Bannon,
Brown, Hooper, and Norman on this topic):

I am still struggling with the dilemma of how to reconcile the role of the designer as both
creator of artifacts that can effect positive change in society and the obverse side of this,
which is where the designer creates artifacts that, though designed with 'good intentions,
3
are totally unsuited to the needs of people.

A recurring debate throughout the book is the view of the computer as a "tool.' Norman
(Chapter 3) makes this one of his themes, that the computer should be viewed as a tool, that it should
serve in the interests of the user, neither dictating the work nor taking control, but rather acting
together with the user to get better mastery of the task domain. Laurel, in Chapter 4, argues against
the notion of computer as tool, but using 'tool' in a different sense, as a intermediary that might stand
in the way between a person and the task or experience. These two points of view are reconciled if the
tool is invisible, transparent, so that the user notices the product and the task, not the tool being used
to get there--a point of view consistent with the arguments for Direct Manipulation in Chapter 5.
Brown comes down strongly in favor of the computer-as-tool view, and he shows that the proper tools
can change one's perception of the task. Indeed, he argues that the best way to induce social change in
a large design community is to provide tools carefully crafted to affect the way designers think of their
task. (Brown uses as an example the design of copiers in his own organization, but the clear intent is
to generalize to design in general, in much larger contexts).

An interesting commentary on the role of computer-based tools in society and how they might
be crafted with concernfor the social implications comes from the Scandinavian experience
in constructing tools for newspaper composition, using high-quality video displays. These
tools are being designed by an innovative design institute, one partially funded by Labor
Unions and Corporations: Norman, Chapter 3, reviews briefly some of their work. (See
Ehn & Kyng, 1984.) A brief overview of the entire project can be found in Howard, 1985.
It turns out that even the intense concern for social impact and explicit support by labor
unions during development did not prevent difficulties when the systems were actually
installed. This emphasizes the point that we should be very cautious in making social com-
mentaries or in introducing technological artifacts designed for a particular social purpose:
We risk tenpering with the complexities of a delicate cultural process with inauficient
knowledge and experience.

3. Those intereted in pursuing these issues might start with the several chaptes in this book that discuss social cbmge, and
might reflect upon Tom Wolfe's (1981) treatment of the leoos from Architecture or upon Giodion's general critique of archi-
tecture (1965).

.' .
 D. A. Norman and D. E. Rumelhart 53 Final Report, NR 667-437

"User Understanding," the theme of Section III of this book, provides a major theme for Brown
in this chapter. He argues strongly that users form mental models of the systems with which they
interact, and he echoes Norman's concern from Chapter 3 that the sstem itself provide the necessary
information to support the development of an effective model (what Norman called the System Image).
Here, Brown builds upon the work in Cognitive Science on understanding (reviewed by Riley, Chapter
7), along with the discussion by Lewis (Chapter 8) of the way that people find explanations of their
experiences. Brown argues strongly that people need to construct explanations of the systems that they
use, that they build upon a "naive theory of computation" (in the sense described by Owen, Chapter
9), and that these play major roles in the way that people recover from difficulties-"repair" the model
and procedures being used. Here he borrows heavily from the experiences of natural language, where
repair and communication take place continually and unobtrusively, points that serve as the themes for
Chapters 14 and 20 by Reichman and Lewis and Norman.

The concept of "Information Flow--the theme of Section VI-is central to Brown's chapter, so
much so that we debated adding this chapter to that section. In many ways that is what the chapter is
about: the flow of information between person and computer system, between people, among social
groups and organizations. Brown takes the strong position that most tasks are interactive, that we
need more tools to support this interaction, and that technology now offers a way to provide these
tools. Design, writing, communicating-all are areas that can benefit from enhanced tools. Tools as
artifact, not only of technology, but of society and culture. The theme that users must help one
another is dominant here, reflecting the arguments made by Owen, O'Malley, and Bannon (Chapters
17, 18, 19, and 21). Brown points out the differences that occur when learning is done in isolation as
opposed to with other people. "Eye contact" is critical for Brown, and he illustrates a case in which
failure to be within sight of others learning a system was detrimental for the learner. As Bannon
points out, in the design of computer (and other) systems, the social organization of work, and even
the physical layout of the office may be as important as the manner by which the system was pro-
gran-,ned.

Brown's chapter-and therefore this review--does not touch upon all of the book. Regardless of
" how you feel about his discussion, we are certain that it will be provocative, raising important issues
about the role of our field within the context of computing, indeed, within the context of society. It
is a fitting way to close the book.

References

Brachman, R. (1978). A structural paradigm for representing knowledge (Tech. Rep. 3605). Cambridge,
MA: Bolt, Beranek, & Newman.
Ehn, P., & Kyng, M. (1984). A tool perspective on design of interactive computer sqppor:for skilled work-

ers. Unpublished manuscript, Swedish Center for Working Life, Stockholm.

Giedion, S. (1965). Space, time, and architecture(4th ed.). Cambridge, MA: Harvard University Press.

*. . . . . . . ... *
D. A. Norman and D. E. Rumelhart 54 Final Report, NR 667437

Goldberg, A. (1984). Smalltalk.S): The interactive programming environment. Reading, MA: Addison.
Wesley.

Howard, R. (1985, April). UTOPIA: Where workers craft new technology. Technology Review. pp. 42.
49.

Miyake, N., & Norman, D. A. (1979). To ask a question, one must know enough to know what is not
known. Journal of Verbal Learning and Verbal Behavior, 18, 357-364.

Wolfe, T. (1981). From Bauham to our house. New York: Washington Square Press.

. . o .
D. A. Normian and D. E. Rumelhart 55 Final Report, NR 667-437

I!

SKILLED MOTOR PERFORMANCE AT VARIABLE RATES

Donald R. Gentner

Motor skills can be performed with a range of overall dwations, and a detailed study of perfor-
mance at different rates provides information about she representation and execution of motor
skills. An analysis of datafrom a variety of motor skills reported in the literature ad a detailed
study of skilled typewriting shows that a central, generalized motor progran with a multiplicative
rate parameter usually does not fit observed performance. A multi-level control model, In which
performance is the composite product of central and peripheralcontrol, provides a better account
of performance data.

Skilled motor performance is based on a combination of innate capabilities and learning in the
environment. Actions such as walking, for which relatively little learning is required to produce com-
petent behavior, lie at one extreme of a continuum. At the other extreme of the continuum are activi-
ties such as playing a violin or flying an airplane, which, although capitalizing on existing motor capa-
bilities, require hundreds of hours of learning to reach expert performance. In all ces, however, flexi-
bility is a striking characteristic of expert performance. Experts are able to modify their actions to
accommodate their intentions and the changing task demands. For example, a basketball player does
not learn to shoot baskets from just a few locations, but instead is able to shoot from anywhere near
the basket, from standing and jumping positions, around defending players, and so forth, Thus, the
skills that a basketball player acquires can be applied in varied an novel situations.

The neural bases of motor skills are usually described in terms of a motor program, but this con-
cept has been given widely differing meanings over the past fifty years. The concept originally was ori-
ginally used action sequences as centrally controlled patterns that functioned with only minor involve-
ment of sensory input, in contrast with the prevailing view of action sequences as stimulus-response
reflex chains (Lashley, 1951; Keele, 1968). The view that motor programs made little use of sensory
input was given support by the finding that animals could make the coordinated rhythmic leg move-
ments typical of locomotion even when their spinal cords and sensory nerves were cut (see Grillner,
1985 for a review). It is now generally recognized, however, that normal motor behavior is based on a
collaboration of perceptual, cognitive and memory processes in the brain, reflexes and pattern genera-
tots in the spinal cord, and sensory input (see, for example, Keele & Summers, 1976; Prinz & Sanders,
1984; Schmidt, Zelaznik, Hawkins, Frank, & Quinn, 1979). For the purposes of this discussion, the

1. The ideas in this paper have developed ova severa year. mnd have benefited greatly from covuerstioos with numerous col-
leaw. I eapecially wish to thank Lmn Abramon, Johs Donald. Mike Jordan, Ronald Knoll. Donald Norman. David
Rumelhaurt, Richard Schmidt, Diane Shapiro. Sad Sterober, Hrn.L~o Teuliap, Paulo Vivini. and Alan Wing.
 D. A. Norman and D. E. Rumelhart 56 Final Report, NR 667-437

concept of a motor program will refer to the centrally stored representations used in the performance
of action sequences.

The simple concept of a central motor program, however, is not sufficient to account for the
observed flexibility of motor performance. It seems implausible that a separate motor program would
be stored for every variation of the action, and there would also be no way to perform novel variations
of an action. The generalized motor program provides a direct account of flexibility in performance.

The Generalized Motor Program

The generalized motor program model was originally proposed by Schmidt (1975) and has since
been discussed by numerous authors (Carter & Shapiro, 1984; Frohlich & Elliott, 1984; Klein, Levy,
and McCabe, 1984; Schmidt, 1982; Shapiro & Schmidt, 1982). A generalized motor program is roughly
analogous to a computer program. Just as a computer program can produce different outputs when it
is invoked with different parameter values, a generalized motor program also has variable parameters
and motor performance will vary depending on the values of the motor program parameters. It is gen-
erally assumed that there is a direct relation between a parameter of the motor program and a feature
of the observed behavior. Evidence for a generalized motor program can be obtained, therefore, by
observing behavior under a variety of conditions, and searching for patterns in the data that can be
related to an underlying parameter in the motor program. An alternative strategy is to search for
invariances in behavior, which would indicate those aspects of a movement that are not the product of
variable parameters.

The most frequently cited candidates for parameters to the generalized motor program are overall
duration (or rate) and force (Schmidt, 1984). For example, Carter and Shapiro (1984) trained subjects
*" to performs a series of four wrist movements with a total duration of about 600 msec. Then they
. asked the subjects to perform the movements as fast as possible. They reported that the overall dura-
* tion for the four movements decreased by about 100 msec for the fast trials, and that the durations of
the individual movements all decreased by the same ratio. The finding that the individual movement
durations maintained a constant ratio was cited as evidence for a generalized motor program with a
multiplicative rate parameter. Similarly, Hollerbach (1981) found that when one subject wrote the
word hell in either large or small script, the timing of vertical accelerations was similar, suggesting that
force of movement could be varied as an independent parameter.

Test for a MultiplIcative Rate Parameter

The most widely cited evidence for a generalized motor program comes from studies of how
action sequences change with changes in overall duration (Schmidt, 1982, p. 311). If the timing of an
action sequence is determined by a generalized motor program with a multiplicative rate parameter,
then the durations of all the components of the sequence should maintain a constant proportion as
the overall duration of the sequence changes. For convenience, I will refer to this model of a general-
ized motor program with a multiplicative rate parameter as the proportional duration model. This
paper is intended to be a critical review of the evidence for a proportional duration model. First, I
present a statistical test for the presence of a multiplicative rate parameter. This test is then used to
examine the evidence for a proportional duration model cited in the literature and to test performance
D. A. Norman and D. E. Rumelhart 57 Final Report, NR 667-437

data collected from expert typists.

A Simple MathematicalModel

To make the argument clearer, consider an action sequence that can be decomposed into a set of
components. For example, the movement of a leg during one walking cycle can be decomposed into a
support component when the foot is touching the ground, and a swing component when the leg is
swinging forward. As the total duration of the step cycle increases, the duration of the support and
swing components will probably change also. The prediction of the proportional duration model is
that, even though all the durations may change, the duration of each component will remain a constant
proportion of the total duration. For example, if the swing component lasts 0.4 s when the total cycle
duration is 1.0 s, the swing component should last 0.6 s when the total cycle duration is 1.5 s. Specifi-
cally, with a multiplicative rate parameter, the ratio d/T should be constant over all instances of the
action sequence, where d, is the duration of the ith component and T is the total duration of the
sequence.

Problems With Existing Tests of Multiplicative Rate Parameter

Although a number of published reports have attempted to examine behavior in relation to a

multiplicative rate parameter, a critical test a critical test of the theory has not emerged thus far.
There are two main weaknesses in published reports.

First, the data examined are often averaged over instances and over subjects. The proportional
duration model describes the timing of individual action sequences. It is true that if the proportional
duration model holds for all the individual instances, it will also hold for the averaged data, but the
reverse case is not necessarily true. The practice of averaging over subjects is especially problematical.
The generalized motor program is necessarily specific to an individual, and data that has been averaged
over subjects cannot be directly related to the model.

Second, in the cases where individual performances were examined, the papers present only a very
limited number of examples. What is needed is a statistical method that permits analysis of individual
sequences, and allows summary of these analyses over many different sequences and subjects.

The Constant ProportionTest

The constant proportion test is a simple and direct test of the proportional duration model. The
basic idea behind the test is that although the total duration for an action sequence may change, the
proportion of time occupied by a given component should remain constant. Therefore, if d, is the
duration of the ith component of an action sequence and 7 is the total duration of the sequence, then
when dll'/ is plotted against T, the corresponding linear regression line should have a slope of zero.
The constant proportion test thus simply consists of determining whether the slope of the linear
regression is significantly different from zero for dlf7 versus T. When a large number of intervals am
being examined, the results of the test can be conveniently summarized over many such linear regre-
sion anal)yses. Specifically, if the criterion of significance is taken to be p< .05, then in a large series of
such analyses we would expect only about 5% of the linear regression slopes to be significantly

." ."" ..... ,.. . : . "".." : ".".'-".":'" . - . "-." . . .. ... ' ..'...- . . " .' ." .. "
D. A. Norman and D. E. Rumelhart 58 Final Report, NR 667437

different from zero if the proportional duration model is valid.

Test Results With Simulated Data

The constant proportion test was verified by using it to analyze simulated data generated accord-
ing to two simple models of timing. Data for the multiplicative rate parameter model were generated
by the equation:

d = r.D, + eL. (1)

Data for the additive rate parameter model were generated by the equation:

di,= Di + r, + et, (2)

In these expressions, d, is the observed duration of the ith component in the nth instance of the
sequence; r, is a normally distributed rate parameter, varying for each instance but constant for all
components within an instance; Di is the mean duration of the ith component; and el, is a normally
distributed random error, differing for each component and instance of the sequence. Each simulated
instance consisted of five intervals, and groups of twenty instances were analyzed with the constant
proportion test.

The results are summarized in Table I. The observed variability in durations is the resultant of
Table I

Analyzing simulated data with the Constant Proportion Test

Proportion of standard deviation

Multiplicative Additive
Rate parameter Random error rejection rate rejection rate

90% 10% 5.0% $0.3%

75% 25% 5.0% 69.2%
50% 6.6% 34.0%
339 67% 7.8% 16.8%
25% 75% 8.2% 12.6%
10% 90% 9.2% 13.4%
0% 100% 11.0% 8.2%
Note. Simulated data were penated for 20 insunce of a S component
sequence. Each rt wa repeated I0 limes.

variability from the changing rate parameter and variability from the random error term. Each value in
the table is based on 100 repetitions of the constant proportion test on a set of 20 instances. Figure 1

,-....
D. A. Norman and D. E. Rumelhart 59 Final Report, NR 667-437

shows one example of the test with data generated by the additive model (Equation 2). Table I shows
how the relative contributions to the observed variability affects the test rejection rate. With the mul-
tiplicative model (Equation 1), the rejection rates are all near the expected level of 5%. With the addi-
tive model (Equation 2), the rejection rates are very high when most of the variability comes from the
rate parameter term, but the rejection rates approach the 5% chance level when most of the variability
comes from the random error term.

There are several technical issues that should be mentioned with respect to the constant propor-
tion test. First, as shown in Table 1, when the variability in durations is entirely due to the random
error term, the rejection rate is approximately 10%, rather than the expected 5%. The increased rejec-
tion rate is caused by the fact that the total duration is the sum of the component durations and the
slope of the linear regression line is not exactly zero. Because the random error term has the same
standard deviation for each interval in the sequence, short intervals will have relatively larger variability
than longer intervals. Thus the relative duration of short intervals will increase with overall duration,
whereas the relative duration of long intervals will decrease. The exact rejection rate obtained with the
constant proportion test depends on the relative mean duration of the component intervals and the
number of instances sampled. The mean durations (100,150,2)0,75,130) and sample size (2D) used for
the results in Table 1 where chosen to be representative of the typing data presented later in this paper.

The second technical issue concerns whether the constant proportion test should be applied to
intervals between components, or to times measured from the start of the sequence. The constant pro-
portion test will work with either intervals or times; the choice should depend on the details of the
timing model being tested. A serial model of timing would generally indicate use of intervals, whereas
a parallel model would generally indicate use of times. The typing data discussed later were tested
using both intervals and times. The rejection rate was slightly higher with times than with intervals.
This probably reflects an earlier finding that typing data fit a serial model of timing better than a paral-
lel model of timing (Gentner, 1982). Also, the relative durations of the intervals in an action sequence
are not completely independent, because they must add up to one, and the slopes from the constant
proportion test must add up to zero.

Literature Review

This section is a critical review of data in the literature that are relevant to the proportional dura-
tion model. Whenever practical, I have analyzed these data with the constant proportion test. The
first section examines laboratory tasks in which the subjects practiced the task for a few hours. Next,
studies of animal and human locomotion are reviewed. The final section takes up other highly prac-
ticed tasks, typically based on thousands of hours of practice.

Laboratory Tasks

Arm movements. Several experimenters have studied arm movements in the laboratory. One of
the original suggestions for a multiplicative rate parameter came from the work of Armstrong (1970).
Armstrong had subjects move a lever with a series of elbow flexions and extensions, while attempting
to match a target pattern of displacement versus time. After four days of practice, subjects became
fairly good at this task, but the interesting finding was that timing errors within an individual trial
D. A. Norman and D. E. Rumelhart 60 Final Report, NR 667-437

0.35
Kt

•,. 0.30

0.5 ,

550 600 650 700 7.10

T

.,--- ".-.
- ".."-''."
- '-""-- ' "-' a -* .i . '-- i ." - " i . " 2- - i','i'i ',* '.-' . .;

Fiipmu I. "[he constanat proportion teat witb date pseated by the additive mdd (P,qston 2). An additive me pmam-
aer anod random ermr contribute equMdy to variability in tbe duration. In ibis cane the dope of the Uno e 8 'paon line was
ignificantly diffc,¢at i, nwzero.
D. A. Norman and D. E. Rumelhart 61 Final Report, NR 667-437

appeared to be related. In particular, Armstrong found that the time from the start of the movement
to the second displacement peak was highly correlated with the time from the start of the movement
to the third displacement peak. The correlation was .88 in the first experiment with a 3 s movement
(Figure 8 in Armstrong, 1970, p 25) and .79 in a second experiment with a 4 s movement (Figure 13 in
Armstrong, 1970, p 34). Of course, we would expect some correlation between these two times,
because the time to reach the third peak includes the time to reach the second peak. I analyzed
Armstrong's data to determine if these positive correlations reflected more than the common time
interval. The first approach was to assume that the time to the third peak represented a total dura-
tion, and use the constant proportion test to determine if the relative time to the second peak is con-
stant. In Armstrong's first experiment, the relative time to the second peak increased from 53% of a
1.8 s total duration to 60% of a 3.3 s total duration (t [for slope * 01 = 2.18, p< .05). In the second
experiment there was no significant change in the relative time to the second peak (t [for slope * 01
-1.83, n.s.).

The second approach to analyzing the data was to compare the time interval from the start of
movement to the second peak (0-2 interval) with the time interval from the second to third peaks (2-3
interval). If the proportional duration model holds, these intervals should be highly correlated within
a trial, and should maintain a constant ratio across trials. These intervals had a significant correlation
(r = .31, p< .05) in the first experiment (Figure 8 in Armstrong, 1970, p25), but their ratio decreased
from 1.07 to .50 as interval 02 changed from .8 to 1.9 s (t [for slope * 01 = -7.5, p< .001). The results
were even more negative for the second experiment (Figure 13 in Armstrong, 1970, p 34); the intervals
had an insignificant negative correlation (r = -.07, n.s.) and their ratio decreased from .61 to .26 as
interval 0-2 changed from 1.4 to 2.7 s (t [for slope * 01 = -8.6, p< .001). In sum, then, the studies of
Armstrong (1970) do not support a proportional duration model. The basic flaw in Armstrong's ana-
lyses was that the correlation Armstrong observed between the times from the start of movement to
the second and third displacement peaks is primarily due to their shared component (the time from
start of movement to the second peak) rather than any consistent change in movement rate. That is, a
and a+b will be positively correlated, even if a and b are completely independent.

This re-analysis of Armstrong's data also illustrates the conservative nature of the constant pro-
portion test. Because the time from the start of movement to the second peak and the time from the
second to third peaks was completely uncorrelated in the second experiment, there was no consistent
trend in the relative time to the second peak, and the constant proportion test did not detect a viola-
tion of the proportional duration model. In mathematical terms, if a and b are uncorrelated dura-
tions, then a/a +b will not change consistently a a function of a +b. Thus the constant proportion
test can only reject data that has a consistent, non-proportional change with overall duration.

Zelaznik, Schmidt, Gielen, and Milich (1995) examined simple horizontal arm movements to a tar-
get, with varying movement distances and movement times. In their first study, subjects were given
about 500 practice trials at each time and distance condition. The movements were 10, 20 and 30 cm
long, with movements times of 150, 200 and 250 ms. Data from individual subjects and instances were
analyzed. Zelaznik el al. (1985) found that although the durations of the positive and negative
accelerations were linearly related to the movement's duration, the times of peak acceleration and
deceleration were not. In fact, they found that the time of peak acceleration was constant and
independent of the movement duration, and same was true of the time of peak deceleration with
 *~~~7 . . . . . . . - 7 -

D. A. Norman and D. E. Rumelhart 62 Final Report, NR 667-437

respect to the end of the movement. A second experiment that focused on the effect of movement
times confirmed these results. Zelaznik el al. (1985) conclude that "the maintenance of relative timing
is not an invariant feature of motor control' (p. 22) in this task.

Wrist twist. Shapiro (1976) had subjects learn a timed sequence of seven wrist twists in 265 trials
over 3 days. A Pattern group attempted to match a given pattern of time and angle targets with a total
movement time of 1600 ms. There was also a Control group that made the same movements attempting
only to match the total time. After completion of the learning trials, there were 15 trials of the move-
ment from memory, and 15 sped-up trials in which subjects were instructed to make the movements as
fast as possible while disreguarding any learned timing pattern. The data were averaged over subjects
and trials. In a com~bined analysis of Pattern and Control groups, Shapiro found no significant differ-
ence between the raemory and sped-up trials, but did find significant differences in the duration pro-
portions for several of the components. The author concluded that the results supported a propor-
tional duration model, but was able to hold this view only by combining the Pattern and Control
groups and proposing that 'the first part of (the] movement sequence is programmed and the second
half is programmed separately" (p. 23). The entire movement sequence does not maintain relative tim-
ing, and this is especially evident when the Pattern and Control groups are analyzed separately.

Carter and Shapiro (1984) reported a similar experiment in which subjects learned a sequence of
four wrist twists. After 600 trials in three days, subjects were asked to ignore the timing they had
learned, and to make the response as rapidly as possible while maintaining spatial accuracy. Data from
the last 10 learning trials were compared with data from the 10 sped-up trials. The data were averaged
over trials and subjects. The average total duration was 570 ms for the final learning trails, compared
with 461 ms for the sped-up trials, and the averaged proportional durations of the components were
not significantly different in the two conditions. However, in addition to the problem with comparing
averaged data, it should be noted that none of the subjects were able learn the original target move-
ment. The target durations varied by a factor of almost 2 (200 msec for the first component and 110
msec for the second component), but subjects made all movements with approximately equal durations
(e.g., 135 ± 16 ms for the first component and 150 t 8 ms for the secotid component). The sped-up
trials also had approximately equal durations for all components. This suggests that the timing was
determined by some simple physical constraint rather than a central motor program, or that subjects
adopted a pattern of equal timing for all movements. In either case, there is little evidence for the
proportional duration model.

Key press. Summers (1975) trained subjects to make a series of nine keypresses with a specified
timing pattern and then asked them to make the keypress sequence as fast as possible while ignoring
the previously learned timing. The training consisted of 473 repetitions of the sequence over 3 days.
Flashing lights above the keys indicated the target timing, which was a repeating sequence: 500-500-100
ms for one group of subjects and 500-100-100 ms for another group of subjects. At the end of training,
there was a test condition in which subjects pressed the same sequence of keys = rapidly as possible
for a total of 110 repetitions. Data were combined across subjects and instances for analysis.
Although subjects were told that maintenance of the previous timing was no longer important or
necessary in the test condition, some remnant of the previous timing was apparent, at least at the
beginning of the test condition. By the end of the test condition, subjects in the 500-500-100 group had
completely lost the previous timing pattern, but subjects in the 500-100-100 group still showed remnants
D. A. Norman and D. E. Rumelhart 76 Final Report, NR 667-437

were three letter words, which did not have separated interstroke intervals, and the interstroke inter-
vals of alternating-hand, three letter words had a strong negative correlation for many typists. For
example, the correlation between the th and he interstroke intervals in the word the ranged from -.67 to
+.54 across the typists, with "- average correlation of -.29. The within-word correlation of .20 for the
separated intervals suggests that about 33% of the observed standard deviation of interstroke intervals
(20% of the variance) was due to changes in typing rate.

In accord with Experiment 1, there was a 31% overall rejection rate in the constant proportion
test. The rejection rate for individual typists ranged from 15% to 50%, once again not supporting the
proportional duration model.

Experiment 3: Random Words

Method. The text for Experiment 3 was a set of 69 words, from 5 to 11 letters in length. The
words were presented in 20 blocks, each block consisting of a different random arrangement of the
words. Presenting the text as random words rather than prose appears to have no effect on the perfor-
mance of skilled typists (Fcndrick, 1937; Shaffer, 1973; West & Sabban, 1982).

Results. The results from Experiment 3 are shown in the third section of Table 4. They are
based on a total of 688 words (4505 interstroke intervals), with an average of 16 instances per word.
The range of durations for a word averaged 33% of the mean duration for that word. The within-word
correlation of .12 for the separated intervals suggests that about 27% of the observed standard devia-
tion of interstroke intervals (12% of the variance) was due to changes in typing rate.

In close accord with the results from Experiments 1 and 2, there was a 36% overall rejection rate
in the constant proportion test. The rejection rate for individual typists ranged from 24% to 48%, all
well above the rate expected with the proportional duration model.

Experiment 4: Slow, Normal, and Fast Typing

One possible objection to the prceeding experiments is that the range of typing speeds was not
wide enough, and any effects of a multiplicative rate parameter were swamped by other effects. Note
that these "other effects" cannot be just random error because, as illustrated in Table 1, even a large
random error with a small multiplicative rate parameter would still produce a rejection rate of approxi-
mately 5%. These "other effects" would thus have to be systematic effects of typing rate on interstroke
intervals. In any event, Experiment 4 was an attempt to increase the relative effect of a rate parameter
by explicitly asking the typists to type at slow, normal or fast rates.

Method. The text was based on a set of 8 sentences (99 words, 611 characters), all in lowercase
and without punctuation. Each page to be transcribed was a different random arrangement of these
sentences. Ten typists were asked to type 6 pages as slow as they could comfortably type, 6 pages at
their normal rate, and 6 pages as fast as possible. The typing rate alternated for succesive pages of
text.
D. A. Norman and D. E. Rumelhart 75 Final Report, NR 667-437

Table 4
Correlations

Typist Wpm Words All Sep % Rejection

Experiment I: Prome

t0] 77 108 0.13 0.20 28

t02 60 109 0.31 0.23 34
t03 71 Joe 0.12 0.13 34
tD4 62 109 0.10 0.13 31
t05 61 l06 0.25 0.18 42
Mean 66 0.18 0.17 34
Ezperiment 2: Promt

tO 90 19 -0.20 0.15 26
tO2 66 19 0.28 0.29 30
t03 76 19 0.04 0.15 50
t4 75 19 0.01 0.28 43
tO5 61 19 0.28 0.33 39
t06 82 19 0.16 0.21 17
t07 110 19 -0.26 0.18 33
to8 112 19 -0.06 0.24 is
t9 86 18 -0.12 0.14 26
:1O 85 19 0.03 020 33
t40 92 18 -0.06 0.03 29
Mean 85 0.01 0.20 31

Experiment 3: Random words

t60 64 68 0.28 0.25 43

t62 101 69 0.02 0.08 31
t62 69 69 0.10 0.11 37
t63 84 69 0.08 0.09 40
t64 92 69 006 0.09 37
t65 68 69 0.14 0.11 37
t66 72 69 0.09 0.08 37
t67 69 69 0.14 014 24
t68 66 69 0.18 0.16 48
t69 66 68 0.06 0.06 26
Mean 75 0.11 0.12 36
Experiment 4: Varied rare

t7O 71 67 0.69 0.67 34

t71 69 67 0.67 0.67 42
t72 71 68 0.64 0.61 33
t73 81 68 0.47 0.49 56
t74 64 68 0.84 0.83 43
t75 101 69 0.94 0.95 72
76 74 69 0.41 0.44 34
t77 69 69 0.92 0.93 35
t78 65 69 0.79 0.79 28
t79 53 69 0.53 0.51 33
Mean 72 0.69 0.69 41

.....-.........
_ "- .. ,.,..,.....,,.,. .. ." .......-..
D. A. Norman and D. E. Rumelhart 74 Final Report, NR 667-437

For each word examined, correlations were calculated between pairs of interstroke intervals across
all instances of the word. All combinations of interstroke intervals were used, but to avoid undue bias
toward long words, a mean correlation was calculated for each word, and these correlations were then
averaged to produce a grand mean correlation for each typist. Parallel models of timing production
can produce negative correlations between adjacent intervals (see Gentner, 1962; Wing, 1980), and could
affect the interpretation of correlations that included adjacent intervals. Therefore, the mean correla-
tion between all non-adjacent intervals was also calculated. As it turned out, neither within-word
averaging nor exclusion of adjacent intervals made any appreciable difference in the analyses.

The constant proportion test. The constant proportion test has already been described. Each
interstroke interval was tested to determine if the relative duration of the interstroke interval remained
constant as a function of the overall word duration. The measure reported is the percentage of tests
that reject a constant proportion with p = .05 or less. Interstroke intervals, rather than times, were
used with the constant proportion test because typing data fits a serial model of timing better than a
parallel model of timing (Gentner, 1982). When the constant proportion tests were repeated with
times since the first letter of the word, the rejection rates were slightly higher.

Experiments I and 2: Prose

Method. In Experiment 1, five typists transcribed a collection of magazine articles adapted from
Reader's Digest totaling 9371 words (55,572 characters) at their normal typing rates. In a similar study
about a year later, the five typists from Experiment 1 and six additional typists participated in Experi-
ment 2, in which they transcribed one of the magazine articles from Experiment 1, with a total of 2D67
words (12,208 characters).

Results. The results from Experiment 1 are shown in the first section of Table 4. They are based
on a total of 542 words (1841 interstroke intervals), with an average of 22 instances per word. On aver-
age, the range of durations for a word was 38% of the mean duration for that word. The within-word
correlation of .17 for separated interstroke intervals suggests that about 31% of the observed standard
deviation of interstroke intervals (17% of the variance) was due to changes in typing rate.

When tested with the constant proportion test, the hypothesis of a constant relative duration was
significantly rejected for 34% of the interstroke intervals. The lowest rejection rate for any typist was
28%. This rejection rate is, of course, much higher than the 5-10% rate expected with the proportional
duration model. It is interesting to note that the rejection rate is also much higher than the 17% rejec-
tion rate found in the simulation studies with a 33% contribution of an additive rate parameter (see
Table 1), indicating a serious deviation from the behavior expected with a simple rate parameter.

Experiment 2 included more typists, but fewer wors than Experiment 1. The results fsom
Experiment 2 are similar to the results from Experiment 1 and ae shown in the second section of
Table 4. They are based on a total of 207 words (584 interstroke intervals), with an average of 18
instances per word. On average, the range of durations for a word wa 49% of the mean duration for
that word. In this experiment, the mean within-word correlation for all interstroke intervals was much
lower than the correlation for only separated intervals. This was because 9 of the 19 words examined
D. A. Norman and D. E. Rumelhart 73 Final Report, NR 667-437

Analyses

The data used for these analyses were the interstroke intervals of words that occurred several
times in the text. In each study, the words used were all words occurring at least ten times in the origi-
nal text, at least 3 letters in length, and consisting of only lowercase letters. The data for a given word
were the interstroke intervals within the word; for example, the four interstroke intervals within a five
letter word. All analyses were based on data from individual typists. Because the analyses can be
strongly affected by outlying data, two procedures were used to eliminate atypical instances of words.
First, instances containing an interstoke interval greater than 400 ms (typically about 9% of the
instances) were eliminated. Second, instances containing an interstroke interval more than 3 standard
deviations away from the mean for that interval (another 5% of the instances) were eliminated.
Finally, if less than five instances of the word remained, the word was not used for that typist.

Correlationalanalysis. When a word is typed repeatedly in the course of transcribing a text, the
interstroke intervals for a given digraph vary from one instance to another. For example, in 20 repeti-
tions of the word system, the interstroke intervals of one typist for the sy digraph varied from 89 to 394
ms (mean = 152, SD = 77). How much of this variation is due to long term variations that affect the
whole word, such as changes in typing rate, and how much is due to short term variations that affect
only a single keystroke? A study of correlations can help answer this question. Assume that the varia-
bility of an interstroke interval is composed of two terms: 1) a common term, common to all inter-
stroke intervals in the word, and 2) an independent term, different for all the interstroke intervals.
The previous question can then be rephrased to ask: What proportion of the observed variability in
interstroke intervals comes from the common term, as opposed to the independent term? If the com-
mon and independent terms are normally distributed, the answer is simple. The proportion of variance
due to the common term is equal to the correlation of two interstroke intervals over instances of the
word. 2 For most purposes, the proportion of standard deviation due to the common term is more
meaningful than the proportion of variance. The relation between correlation and standard deviation
is more complicated. If p is the proportion of standard deviation due to the common term, then the
correlation coefficient is given by the following sigmoidal function.

r = __ p2 (3)
P2 + (1-p)

Even if this simple model with normally distributed common and independent terms is not completely
valid, a correlational analysis can still provide an approximate measure of the proportion of variability
that is constant within an instance of a word.

2. In - earlier parei (Gentner. 192) 1 mated that an mcn correlation between intervals of .2 indicated that rate cbmis e-
counted tof 4% (r') of the variance. This w an error. A cordation of 2 muawms that 2D% of the variance is due to fate
cbange
D. A. Norman and D. E. Rumelhart 72 Final Report, NR 667-437

Music. Music is different from the previous tasks because, with music, the control of timing is
an explicit part of the performance. Musical notation directly specifies the timing structure of the
piece, although performers may deviate from regular timing for expressive effect (Shaffer, 1984).

Michon (1974) analyzed tape recordings of a performance of Vexation by Erik Satie for evidence
of changes in the temporal pattern with changes in overall duration. Vemaions is aptly named, because
the performer must repeat a five part motif 840 times. The entire performance took about 19 hours.
The duration of the individual parts varied roughly between 15 and 25 s, and Michon found that this
duration was the major factor accounting for deviations from the prescribed timing. That is, the tem-
poral structure of the performance varied with the rate of playing. However, Michon did not present
any quantitative evidence for this conclusion.

Clarke (1982) also examined performances of Vexationi by Erik Satie. Two graduate musicians
each performed about an hour of the piece on the beautifully instrumented piano in Shaffer's labora-
tory. Clarke divided up the repetitions into six groups based on overall duration of each part, and
found significant effects of part duration. In particular, the faster instances (about 13 s duration) tend
to become slower as they proceed, and the faster instances (about 22 s duration) tend to become
slower. The constant proportion test did not show any significant changes in relative duration for the
quarter note intervals from subject M (Figure 3A in Clarke, 1982, p 8), but 6 of 11 quarter note inter-
vals showed a significant change in relative duration for subject G (Figure 3B in Clarke, 1982, p 8).

Typewriting Data

In addition to analyzing data from the literature, I have applied the constant proportion test and
other analyses to a extensive body of data that I have collected over the past five years from expert typ-
ists.

General Method

All subjects were professional typists recruited from the university and local businesses. They
typically typed 14 hours per week (range = 2 to 25), and their median typing speed was 71 words per
minute (range = 53 to 112) measured over approximately 40 minutes of typing and uncorrected for
errors. Subjects were paid for participation. The five subjects in Experiment 1 were also subjects in
Experiment 2. Otherwise, subjects participated in only one experiment.

In Experiment 1, subjects typed at a Hazeltine 1500 computer terminal connected to a minicom-

puter that recorded keypresses and the corresponding times. Typed characters were displayed on the
terminal screen. The typists were all familiar with this terminal, having used it in conjunction with the
campus word-processing system. In Experiments 2 through 4, subjects typed on a high quality, elec-
tronic keyboard (Microswitch 51SD12-4) with a layout identical to the IBM Selectric typewriter and a
similar 'feel'. The keyboard was connected to a microcomputer that recorded keyprtsses and the
corresponding times with an accuracy of 1 ms. Typed characters were displayed on a CRT screen in
front the the typist. The text to be transcribed was presented as printed or typewritten copy in a con-
venient place next to the keyboard. After a warmup period, typists were asked to transcribe the text at
their normal, rapid rate and to ignore any errors they might make.

-.
. .
- • . - . , . . . . . • • • . . - . . ,- . .. -
I' a . .. .. " . ' " -- '- " - " : " -i/ l M F dl IMtlIIM III. l . llll. . l~m --.A t M ~
D. A. Norman and D. E. Rumelhart 71 Final Report, NR 667-437

Table 3

Interstroke Intervals and Times in Typewriting

Unit % at short duration % at Ions duration t (slope * 0) Significance

neloud (42 instances; total duration - 843 to 1177 ms)

Interstroke intervals

en 10.8 11.3 .69 n.s

1C 19.1 20.5 1.19 1.
ci 9.3 13.0 5.66 < .00
lo 18.5 14.7 .6.57 <.001
o 10.8 8.7 .5.44 < .001
sc 10.7 14.2 5.64 < .001
ed 20.8 17.7 .5.37 <.001
Time ince first letter

a 10.8 11.3 .69 e.g.

c 29.9 31.7 2.62 <.02
1 39.2 44.7 6.33 <.001
o 57.7 59.4 2.26 <.06
* 68.5 6.1 -.62 u.s
79.2 82.3 5.37 <.001
troube (27 instances; total duration - 545 to 1218 ms)

Interstroke intervals

It 21.6 18.5 -3.57 < .02

to 14.1 19.2 4.83 <.001
ou 17.0 13.9 .2.54 <.02
ub 14.7 18.7 2.74 <.02
bl 17.4 14.9 .1.92 u.s.
le 15.1 14.8 -. 37 e.s.
Tame mace first letter

r 21.6 18.5 .3.57 < .002

o 35.5 37.7 1.85 1s.
u 52.8 51.6 .87 U.S.
b 67.5 70.3 2.72 < .02
I 84.9 35.2 .37 1.s.
Note. Dats for the word eac od is from Te'nuolo and Vivimi (1930).
Data for the word tr=Wde is from Ternuolo ad Viviani (1979).

repeated instances of the word. (This prediction is closely related to the constant proportion test used
in the present paper, but I now feel that the constant proportion test is a more direct test of the pro-
portional duration model than the test for a constant ratio of intervals.) The constant ratio analysis
was restricted to the pairs of intervals with significant positiv correlations, but even then. 60% of the
pairs deviated significantly from a constant ratio. Te results from both tests were taken as evidence
against the proportional duration model.
D. A. Norman and D. E. Rumelhart 70 Final Report, NR 667-437

In summary, the studies of handwriting have yielded mixed results. The data of Hollerbach (1981)
and Wing (1978) do not fit the proportional duration model, whereas the data of Viviani and Terzuolo
(1980) support the proportional duration model.

Typing. In addition to the study of rapid speech mention earlier, Sternberg, Knoll, Monsell, and
Wright (1983) also examined the rapid typing of short letter sequences. Like their finding with short
speech sequences, the interstroke intervals of short typewriting sequences increase as the length of the
sequence gets longer. In order to determine whether these increases in the interstroke interval were
evenly distributed over the interval, Sternberg el al. examined the index finger trajectories in
alternating-hand sequences. They divided the trajectory into five components, and found that the
increase in the interstroke interval (actually it was two interstroke intervals, because they examined an
alternating hand sequence) was entirely localized in one component of the trajectory. The 'late lift'
component increased from 80 ms (36% of the trajectory) to 240 ms (62% of the trajectory), as the
number of letters in the sequence increased from 3 to 5 and the total trajectory duration increased
from 220 to 390 ms. When tested with the constant proportion test, the relative duration changed sig-
nificantly for 3 of the 5 components. The results are very similar to their results from short speech
sequences: the changes were localized to one component of the sequence, and therefore to not fit the
proportional duration model.

Probably the most commonly cited evidence for the rate parameter model is the work of Terzu-
olo and Viviani (1979, 1980; Viviani and Terzuolo, 1980) on typing. Terzuolo and Viviani examined the
performance of expert typists during continuous typing to determine how the keystroke times within a
word change as the overall duration of the word changes. Terzuolo and Viviani found that when the
keystroke times for repeated instances of a word are plotted in the proper manner, the times for each
letter appear to radiate from a common origin, thus showing that relative timing was maintained as the
overall duration of the word changed. Terzuolo and Viviani have presented these radial plots for two
words: 42 instances of the word enclosed (Figure 6A in Terzuolo and Viviani, 1980, p 1092; and Figure
2A in Viviani and Terzuolo, 1980, p 527) and 27 instances of the word trouble (Figure IA in Terzuolo
and Viviani, 1979, p 115). Although these radial plots are intuitively clear and appealing, there is no
quantitative measure to determine how well the data fit the radial lines, or to summarize results over a
large number of words and typists. In fact, as shown in Table 3, when analyzed with the constant pro-
portion test these data do not support the proportional duration model. For both words, the majority
of intcrstroke intervals did not maintain the same relative duration as the overall word duration
changes. Similar results were obtained when the data for the two words was analyzed in terms of time
since the first letter rather that in terms of interstroke intervals. Thus, the typing data of Terzuolo
and Viviani do not support the proportional duration model.

In an earlier paper (Gentner, 1982), 1 examined data from prose transcription by a group of expert
typists for evidence of the proportional duration model. Two predictions of the proportional duration
model were tested on data from repeated words in the text. The first prediction was that two different
interstroke intervals within a word should be positively correlated over repeated instances of the word.
For example, if the wi interval in with is longer than average for a particular instance of the word, the
th interval in that instance should tend to be longer also. Of the 1,517 pairs of intervals examined, only
15% had significnntly positive correlations. The proportional duration model makes a second, stronger
prediction: that two interstroke intervals within a word should tend to maintain a constant ratio over
D. A. Norman and D. E. Rumelhart 69 Final Report, NR 667-437

In other analyses of these EMG data, Tuller, Kelso, and Harris (1981) found that "no measured
interval was found to vary systematically with changes in speaking rate' (p. 73), and that "the temporal
relationships between these muscles were not systematically affected by changes in speaking rate" (p.
73).

In a related study, Tuller, Harris, and Kelso (1981) found that, contrary to the proportional dura-
tion model, overlap between articulator muscle activities was independent of acoustic syllable duration
(Figure 4 in Tuller el al., 1981, p 48). The proportional duration model predicts that all aspects of
speech production should scale proportionally in time. The results of Tuller, Kelso, and Harris (1982)
and Tuller, Harris, and Kelso (1981) are consistent with the finding by Kozhevnikov and Chistovich
(1965) that speech events at the level of the syllable do not correspond to the proportional duration
model.

Sternberg, Monsell, Knoll, and Wright (1978) have found that when subjects perform a rapid
action sequence, the rate of performance is dependent upon the number of units in the sequence. For
example, the time to speak a digit is about 8 ms longer in a sequence of 5 digits than in a sequence of 4
digits. Sternberg, Knoll, Monsell, and Wright (1983) inquired whether this change in speaking rate was
equally distributed through the word or was localized to a particular portion of the word. They had
subjects rapidly speak sequences of two-syllable words, such as copper and token. The data were com-
bined over instances, words, and subjects. When the acoustic waveform for each word was decom-
posed into six components, the increase in duration was almost entirely a result of an increase in dura-
tion of the second vowel, which increased from 17 ms (8% of the word) to 42 ms (17% of the word), as
the number of words in the sequence increased from 2 to 5 and the total word duration increased from
216 to 249 ms. When the speech data (Figure 25 in Sternberg et al., 1983, p. 16d) were tested with the
constant proportion test, the relative duration changed significantly for 4 of the 6 word components.

Handwriting. Hollerbach (1981) reports vertical acceleration profiles for the pen of a subject
writing the word hell at three different rates (Figure 28 in Hollerbach, 1981, p 154). If the letter boun-
dary is taken to be the moment of zero acceleration at the bottom of the stroke, then the relative
durations for three of the 4 letters changed significantly as the word duration changed from about 800
ms to about 2000 ms (all values of t [for slope * 01 > 6.6, p< .05).

Viviani and Terzuolo (1980) asked people to write letters of the alphabet at different speeds and
recorded the pen positions with a digitizing table. They report velocity profiles for nine instances of a
letter written by one person (Figure 4 in Viviani & Terzuolo, 1960, p 529). Analysis of their data with
the constant proportion test indicated that none of the intervals between velocity maxima and minima
showed a relative change as the letter duration changed from 245 to 680 ms (first maxima to last max-
ima), thus supporting the proportional duration model.

Wing (1978) examined timing within single handwritten letters (v, n, w, and m). Based on corre-
lations of individual instances, he concluded that "these data do not show overall time scaling of suc-
cesive segments" (p. 164).
 D. A. Norman and D. E. Rumelhart 68 Final Report, NR 667-437

Other Highly PracticedSkills

This final section of the review focuses on other highly practiced skills. Some are innate actions,
such as breathing; others are learned skills in which the subjects have had thousands of hours of prac-
tice, such as piano playing.

Breathing. Clark and von Euler (1972) studied the relationships between depth of a breath and
the duration of the inspiratory and expiratory p --;es in cats and humans. They found an approxi-
mately linear relation between the inspiration and expiration durations. Clark and von Euler also state
that the durations were approximately proportional for humans, but the durations were not propor-
tional for cats. Analyses of data from individual breaths for one human and one cat with the constant
proportion test confirmed these findings. For the human (Figure 9a in Clark & von Euler, 1972, p.
286), the relative duration of the inspiration phase remained at 47% as the total duration changed from
1.2 to 3.7 s (t [for slope * 0] = -.04, n.s.). In contrast, for the cat (Figure 9b in Clark & von Euler,
1972, p. 286), the relative duration of the inspiration phase increased from 38% to 50% as the total
duration changed from 1.4 to 2.4 s (t[for slope * 0] = 18.1, p< .001).

Speech. Kozhevnikov and Chistovich (1965) studied how the components of speech vary as the
overall speed changes. They had subjects repeat phrases with overall durations ranging from 1 to 3 s,
and determined the duration of various components from measurements of speech articulator move-
ment. They examined the relative duration of words, syllables, and vowels and consonants. The data
were averaged over instances and subjects. When analyzed with the constant proportion test, the rela-
tive duration of words (Figure 3.5 in Kozhevnikov and Chistovich, 1965, p. 83) changed significantly in
2 of 9 cases. This rejection rate is within chance (binomial distribution, p = .23) The relative duration
of syllables in the word topila (Figure 3.7 in Kozhevnikov and Chistovich, 1965, p. 86) did not change
significantly in any of 9 cases. In contrast to behavior of words and syllables, however, the relative
duration of consonants (Figure 3.9 in Kozhevnikov and Chistovich, 1965, p. 88) decreased from 36% for
a total phrase duration of 800 ms to 29% with a total phrase duration of 3000 ms (t[for slope * 01 -
-12.7 p< .001). This finding that the relative duration of consonants and vowels change as the rate of
speaking changes is commonly reported. Thus the constant proportion test does not describe speech at
the syllable level, but does provide a good mode of longer units of speech.

Tuller, Kelso, and Harris (1982) recorded EMG patterns associated with tongue, lip and jaw
movements while words were spoken with different speeds and stress. They searched for linear rea-
IL tionships between a component of the EMG pattern and a longer period of activity by calculating
correlations for onset, offset and peak amplitude of EMO for muscles associated with vowel and con-
sonant production. Data were averaged over instances before analysis. Eight of the nine relationships
reported showed a wide variety of positive and negative correlations. Only one relationship showed a
. consistently high correlation: in the sequence VICV2, the time from onset of V, to onset of the con-
sonant C was highly correlated with the time between onset of V 1 and V 2 . Although a high correlation
indicates a linear relation, but not necessarily a proportional relation, examples of single-subject data
for the one cae with high correlations (Figure 4 in Tuiler el al., 1982, p 467) were accepted by the con-
stant proportion test in 11 of 12 cases.

• . . ... . ..... .: -..: .I: . ,.. . ,. .: , ., . -.

 D. A. Norman and D. E. Rumelhart 67 Final Report, NR 667-437

each speed (Figures 2A and 2B in Grillner el al., 1979, p 179). The constant proportion test could not
!Pbe applied because data were plotted only for slowest and fastest running and walkin speeds, but I did
compare these extremes. The support phase occupied an averge of 58% of a 915 ms cycle at the
slowest walking speed, and 49% of a 630 ms cycle at the fastest walking speed. The within-subject
difference in proportion was significant (t [for slope * 01 = 4.1, p <.001). Similarly, the support
phase occupied an average of 45% of a 725 ms cycle at the slowest running speed, and 25% of a 530 ma
cycle at the fastest running speed. The within-subject difference in proportion was significant (t [for
slope # 01 -8.6, p < .001).

Grillner el al. (179) also present data for two of their eight subjects on the duration of the fiex-
ion phase at the knee joint during running and walking. The data are for individual subjects and step
cycles. The duration of the flexion phase was linearly related to the step cycle duration, but it was not
Sproportional to step cycle duration; that is, when flexion or extension duration is plotted against step
cycle duration, the fitted straight line cannot be extrapolated through the origin. These conclusions
were confirmed when tht; data were analysed with the constant proportion test. For one subject (Fig-
ure IA in Grillner et al., 1979, p 178) the relative duration of the flexion phase decreased significantly
as the walking step cycle increased from 620 ms to 918 m (tffor slope * 0) = -3.56, p< .01), and the
relative duration of the flexion phase also decreased significantly as the running step cycle increased
from 519 ms to 758 ms (t [for slope * 01 = -2.73, p< .02). For a second subject, one of two with dif-
ferent slopes for running and walking when flexion duration was plotted against step cycle (Figure IB
in Grillner et al., 1979, p 178), the relative duration of the flexion phase did not change significantly as
the walking step cycle increased from 553 ms to 897 m (t[for slope * 01 = 1.24, n.s.), whereas the rela-
tive duration of the flexion phase decreased significantly as the running step cycle increased from 482
. ms to 705 ms (t[for slope * 01 = .5.14, p< .001).

Lee, Lishman, and Thomson (1962) measured the strides of skilled long jumpers during their run-
up to the takeoff board. They were primarily interested in how the jumpers adjusted their gait to land
just in front of the takeoff board for the jump, but they report data from the individual step cycles
that was analyzed with the constant proportion test. The data reported were for individual jumpers
for each stride before takeoff, but averaged over 12 runs (Figure 4 in Lee et al., 1962, p 453). Data for
the last stride before takeoff was omitted from these analyses because it was atypical. Subject MN's
support phase duration varied from 47% of a 218 ms step cycle to 60% of a 272 ms step cycle (t[for
slope * 01 = 5.0, p< .001). Subject VW's support phase duration varied from 48% of a 225 ms step
cycle to 63% of a 281 ms step cycle (t [for slope # 01 = 4.7, pC .I01). Subject FM's support phase
duration varied from 50% of a 220 ms step cycle to 60% of a 246 ms step cycle (i [for slope * 01 - 2.0,
n.s.). Thus, the step cycles of two of three jumpers are not consistent with the proportional duration
model.

Winter (1983) reports ankle, knee and hip joint angle curves for people walking at their natural
pace (571 ms mean step cycle), as well as other groups walking at a slow pace (708 ms) or a fast pace
(493 ms). The data were averaged over 14 to 16 subjects within each pace group and over an unstated
number of step cycles for each subject. Plots of the avcrae joint angles against the normalized stride
cycles appear remarkably similar for the three pace groups, but there is insufficient information for any
statistical test of the proportional duration model.

S-A ............................
 D. A. Norman and D. E. Rumelhart 66 Final Report, NR 667-437

Table 2

Durmio of Step Cyde nmi in the Cat

Phbue % a Short Duratio %at Lon DaIvtioo t (dope 0 0) Sipificance

Trot (12 instances, total dratio. - 345 to 0 ma)

El 22 16 -1.51 a.
E2 19 12 -2.74 <.02
3 27 53 4.0 <.001
F 31 18 -3.24 <.01
Gallop (A insancms total duration - 253 to 382 mu)

El 28 29 .10 1.1.
E2 14 13 -. 30 0.s.
E3 18 18 .13 U.S.
F 39 41 .15 e.g.
Note Data an from Golow. Ra nkinS, and Stuart (1973).

Human. Shapiro, Zernicke, Gregor, and Diestel (1981) filmed experienced runners while running
and walking on a motor driven treadmill. They measured the duration of the four phases of the Phil-
lippson step cycle at four walking speeds and four running speeds. Data were combined over subjects
and instances before analysis. The authors report that there was a significant effect of speed on the
proportion of the step cycle occupied by the components, but most of the change was between pits,
and they did not find a significant difference in relative durations within the walking or running pits.
1These results are not directly relevant to the rate parameter model, however, because Shapiro et al.
- (1981) examined proportions as a function of speed of locomotion, rather than as a function of total
step cycle duration. In fact, although the average step cycle durations for walking have a reasonable
range of about 33%, the range of step cycle durations for running is only 4% and 3 of the 4 durations
are within 1% (see Figure 3A in their paper), and thus the running data do not provide information
- about how the components of the step cycle change with changes in the total duration. When the
data for walking (from Figures 3 and 4 in Shapiro et al., 1981, pp. 41-42) were analyzed with the con-
stant proportion test, the mean relative duration of three of the components of the Phillippson step
cycle did not change significantly with total cycle duration, but the fourth component, E3, increased
from 48% of a 954 ms cycle to 54% of a 1320 ms cycle (i [for slope * 0] = 6.9, p< .01). When the walk-
ing data (Figure 3 in Shapiro et al., 1981, p. 41) were analyzed in terms of swing and support phases,
there was no change in the relative duration of the support phase (i [for slope * 0] - 2.4, n.s.). The
net results, then are somewhat mixed. With one exception (the E3 phase in walking), the analyses sup.
port the proportional duration model, but the main problems remain: the data were combined across
subjects and the running data do not span a significant range of cycle durations.

Grillner, Halbertsma, Nilsson, and Thorstenason (1979) studied humans walking and running on a
treadmill at velocities from 1.7 to 8.3 m/s. They recorded joint movement with a knee goniometer, and
mounted foot switches on the subjects' shoes to record the duration of foot contact with the ground.
Cycle duration data are reported for seven individual subjects, but are averaged over ten instances at

s w "
 D. A. Norman and D. E. Rumelhart 65 Final Report, NR 667-437

Delcomyn found that, on average, the support phase occupied about 51% of the step cycle when the
cycle time was 50 msec, but increased to 71% of the cycle when the cycle time was 400 ms. The data
presented in Delcomyn's (1971) article are for individual instances of step cycles; it is unclear whether
the data is for one or several cockroaches. I analysed data for individual instances of step cycles (from
Figures 5A and 5B in Delcomyn, 1971, p449) with the constant proportion test. The support phase
duration of leg R 3 increased from 47% of a 52 ms step cycle to 70% at a 440 ms step cycle (t[for slope
S0] = 8.3, p< .001). The support phase duration of leg R 2 increased from 51% of a 53 ms step cycle to
" 78% at a 465 ms step cycle (: [for slope * 01 = 9.5, p< .001). These analyses support Delcomyn's con-
clusion that "both forward [swing] and rearward [support] movements of the legs relative to the body
"-" decreased in duration as the insect's rate of forward progression increased, but at different rates" (p.
452).

Lobster. Macmillan examined walking in the American lobster, recording movements on video-
tape and also recording electromyograms from selected muscles. Two of Macmillan's figures report
data from individual instances of movement that I was able to analyze with the constant proportion
test. Both sets of data were consistent with the proportional duration model. The proportional dura-
tion of the swing phase measured from the videotape record (Figure 21a in Macmillan, 1975, p 31) did
not change significantly as the step cycle time increased from 1000 ms to 1800 ms (i [for slope * 0] =
- 1.31). Similarly, the proportional duration of the depressor bursts measured from the electromyograph
record (Figure 29 in Macmillan, 1975, p 31) did not change significantly as the step cycle time increased
from 600 ms to 120D ms (t[for slope * 01 = -. 18).

Tortoise. Williams (1977) analyzed motion picture films of walking tortoises. She reports that "a
change in the forward speed of locomotion was accomplished by an approimately proportionate
change in the duration of all portions of the cycle' (p. 54). No data or statistical analyses are reported.

Cat. Goslow, Reinking, and Stuart (1973) analyzed high-speed motion pictures of cats moving
freely at speeds from 1 to 16 miles/hr (.4 to 7 m/s). Cats use three gaits over these speeds: walk (up to
.7 m/s), trot (.7 to 2.7 m/s) and gallop (above 2.7 m/s). Across this range of speeds and gaits, Goslow el
al. (1973) found that the duration of the support phase remained approximately constant at 20 ms,
whereas the duration of the swing phase decreased from about 500 ms at the slowest speeds to less than
100 ms at the highest speeds (Figure 3 in Goslow et al., 1973, p 12). They also reported the durations
of the four phases of the Phillippson step cycle (Figure 4 in Goslow et al., 1973, p 13). The data were
for individual instances, combined over nine cats. There were insufficient data from the walking gait,
but it was possible to analyze the data from the trot and gallop gaits with the constant proportion test.
These analyses are summarized in Table 2. The proportional duration model does not fit Goslow et
al.'s data for the trot gait, but the proportional durations of the step cycle phases are remarkably con-
stant for the gallop gait.

Dog. Arshavskii, Kots, Orlovskii, Rodionov, and Shik (1965) measured the leg joint angles of
dogs running on a treadmill. The joint angle records were used to calculate the durations of the sup-
port and swing phases of the step cycle (Figure 3A in Arshavskii et al., 1965, p 741). The data
presented were individual instances from one dog. Analysis with the constant proportion test did not
support the proportional duration model. The support phase duration increased from 45% of a 420 ms
step cycle to 69% of a 765 ms step cycle (t[for slope * 0] = 5.9, p< .001).
 D. A. Norman and D. E. Rumelhart 64 Final Report, NR 667-437

1200

1000

7500

.00

* 7500 *

CO

05

0*

250D 7010015

Cycle time (mSec)

Figure 2. Vaiation in the bunst duration of motor ason Da with cycle time for a cockroach. The ded line is an ca-
* mpie of bow the data might look if the depresor duation maintained a constant proportio, of the total cycle dime. The
difference between the line through the data points and the diagonal line gives an approximate messure of the bum duration
of levatoi motor non 5. Note. From 'Central progiramming ad reflex control of walking inthe cockroacV by K. 0. ftauson.
IM7. Jowal of' EiriamemI iology, 56. p 182. Copyright 197 by The Company of Biologiss Limited. Reprinted by permit-
@ion.
 D. A. Norman and D. E. Rumelhart 63 Final Report, NR 667-437

of the previous timing pattern. Although Summer's (1975) results are often cited in support of the
proportional duration model, the results actually strongly contradict such a model. The ratio of slow
to fast interkeypress intervals in the final block of training for the 500-500-100 groups was 579/252 = 2.3,
but the corresponding ratio in the test block was 292/246 = 1.2. Similarly, the ratios for the 500-100-100
• .group was 421184 = 2.3 in the final block of training and 283/185 = 1.5 in the test block. It appears
that the fast intervals was at a floor, and the only effect of speeding up is to speed up the slow inter-
vals. Proportional durations were not preserved in the test condition, as required by the proportional
S.-duration model.

Glencross's (1973) study of subjects turning a hand crank has been cited by several authors (e.g.
Shapiro, 1976; Zelaznik et al., 1985) as providing support for the proportional duration model. How-
ever, this study compared performance across subjects, each of whom turned the crank at a single
speed, and therefore the results are not relevant to the rate parameter mo Iel, which describes how an
individual performs at differing rates.

Locomotion

In contrast to motor tasks practiced for a few hours, locomotion is largely innate and has been
studied in a wide variety of species.

Cockroach. Most of the studies in the literature report behavioral data because it is the easiest
data to collect. A few investigators, however, have examined data from neural recordings. Pearson
(1972) presents an impressive study based on single motor neuron recordings in cockroaches. Pearson
measured the burst durations in levator and depressor motor axons. Figure 2 shows the burst duration
of the depressor motor axon from a remarkable cockroach that varied its total step cycle from 84 ms to
1265 ms. The observed depressor burst durations clearly do not maintain a constant proportion of the
total cycle duration. The constant proportion test indicated that the proportion of the total cycle
occupied by the depressor bursts increased significantly as the total duration increased (t [for slope
01 = 22.3, p< .001). The depressor bursts occupy 42% of the step cycle when the cycle time is 84 ms,
but this proportion increaises to 87% of the step cycle when the cycle time is 1265 ms. In fact, for cycle
* times g -ater than 400 ms, changes in the depressor durations account for almost the entire increase in
cycle time, and the duration of the levator bursts remains approximately fixed.

When the constant proportion test was applied to data from another cockroach (Figure 5 in Pear-
son, 1972, p 181), none of the durations showed evidence of t multiplicative rate parameter. The dura-
tion of levator axon 5 decreased from 46% of the step cycle at a 150 ms step cycle to 34% at a 500 m
step cycle (t [for slope * 0] = -7.8, p< .001). The duration of levator axon 6 decreased from 29% of the
step cycle at a 86 ms step cycle to 19% at a 500 ms step cycle (t [for slope * 01 = -6.8, p< .001). And
the duration of depressor axon D, increased from 40% of the step cycle at a 98 ms step cycle to 68% at
* a 530 ms step cycle (t [for slope * 0] = -14.4, p< .001).

Pearson's observations of motor neuron activity correspond nicely to the behavioral observations
of cockroach walking by Delcomyn (1971). Delcomyn recorded the walking movements of cockroaches
with a high speed motion picture camera. He found that cockroaches used a single, alternating triangle
pit at all speeds of locomotion from 5 to 80 cm/s, corresponding to step cycles of 400 to 45 ms.

";.', ,. *-. ." '.. ..- . . . . . :' / .,-.' . ..- .' .",.".... . .. . ." . * *, .-.. - *. ., ,
D. A. Norman .ad D. E. Rumelha[ 77 Final Report, NR 667-437

Results. The typists were able to vary their rate on demand. Table 5 lists their typing rates for

Table 5

Typing Rate in Ezpe-imet 4

"ypia Slow Normai Fast

070 44.2 73.5 80.8

C71 49.0 6.0 73.8
t72 48.4 70.9 73.5
t73 61.7 M1.0 83.9
t74 38.4 61.9 86.1
t75 34.1 10.5 105.4
t76 60.7 74.2 @0.3
t77 25.1 68.5 75.3
t7a 37.8 65.3 6.8
t79 37.2 53.3 56.S

MEAN 43.7 72.0 79.2

the three experimental conditions. Not surprisingly, typists were more effective at slowing down from
their normal rate than at speeding up. For these analyses, data from all three conditions were pooled;
the three conditions just served as a device to encourage a wide range of word durations. The cutoff
point for eliminating outlying instances was increased from the normal 400 ms to an interstroke interval
of 700 ms because of the many long interstroke intervals in the slow condition.

The results from Experiment 4 are shown in the last section of Table 4. They are based on i total
of 685 words (4505 interstroke intervals), with an average of 18 instances per word. The range of dura-
tions for a word was much larger than in the previous experiments: an average of 107% of the mean
duration for the word. The within-word correlations were also much larger, ranging from .44 to .95 for
separated interstroke intervals, reflecting the larger changes in typing rate. The average within-word
correlation of .69 for the separated intervals suggests that about 60% of the observed standard devia-
tion of intcrstroke intervals (69% of the variance) was due to changes in typing rate.

Once again, the hypothesis of a constant relative interstroke interval had a high rejection rate; in
this experiment the mean rejection rate was 41%. The rejection rate for individual typists ranged from
28% to 72%. Thus, increasing the range of typing rates actually led to a stronger rejection of the pro-
portional duration model.

Patternsof Change With Typing Raie

If, as has been demonstrated, the interstroke intervals of skilled typists are not described by a
proportional duration model, is there any systematic pattern of rate change, or am these data merely
the result of random variation? There are :hree strong arguments indicating that expert typists vary
their rate. First of all, Experiment 4 clearly demonstrates that typists can vary their typing rate over a
large range on demand. Second, the crrelational analyses show that even when typing at normal rate,
fluctuations in rate occur that persist for at least the duration of a word. Third, an autocorrelation
study indicated that while typing at their normal rate, some typists have rate changes that persist for
 D. A. Norman and D. E. Rumelhart 78 Final Report, NR 667-437

many words. The word durations from Experiment 3 were normalized by dividing the observed dura-
tion by the mean duration for that word and typist, and then the autocorrel.tion coefficient was deter-
mined for lags of I to 50 words. The extent of rate changes was conservatively estimated by the lag
before the first negative autocorrelation. Individual differences between typists were very pronounced.
Rate changes persisted for only 1 word for two of the typists; 3 to 6 words for two typists; 13.31 words
for three typists; and more than 50 words for the remaining three typists. So, for at least some typists,
there are normally slow changes in typing rate that persist for many words.

Given that typing rates can change, are there any regularities in the details of how performance
changes with overall rate? David Rumelhart and I analyzed the intertroke intervals of typists who
transcribed normal prose under fast and slow conditions. In the fast condition, they were told to
assume they were typing a rough draft and to type as fast as possible, without concern for errors. In
the slow condition, the typists were told to assume they were typing a final copy and to type carefully
and minimize errors. For purposes of analysis, the letter-letter digraphs were divided into four classes
based on the fingers used in touch typing: doubles, such as dd; 1-finger digraphs, non-doubles typed by
a single finger, such as de; 2-finger digraphs, typed by different fingers on the same hand, such as de;
and 2-hand digraphs, typed by different hands, such as de. The results of the analysis are shown in
Table 6. Interstrokc intervals for 2-hand digraphs had the largest relative increase, followed by 2-finger

Table 6

% Increase in ISI (slow-fam)

Wpm Wpm
Typist fag slow Double Ifinget 2fingr 2hand

t02 72.5 591 7.6 6.7 2D.0 28.5

t03 8D.3 72.9 1.5 5.4 11.7 12.9
tO4 94.3 73.1 4.1 9.7 9.2 16.8

MEAN 79.0 68.4 4.4 7.3 13.6 19.5

and 1-finger and double digraphs, in descending order. The mean absolute change in interstroke inter-
vals followed the same pattern. This, of course, is in direct opposition to the proportional duration
model, which predicts that because the interstroke intervals of 2-hand digraphs are shortest, they
should show the least absolute change in duration. Instead, 2-hand digraphs (mean = 109 ms) increased
by an average of 22 ms, whereas doubles (mean = 185 ms) increased by an average of only 7 ms. An
analysis of variance confirmed that digraph class had a significant effect, F(3,6) = 8.57, p=.014.

The same pattern of change was found when data from the normal and fast conditions in Experi-
ment 4 were compared. The mean percent decrease in median interstroke intervals was 3.9, 6.8, 10.0,
and 12.3 for doubles, 1-finger, 2-finger, and 2-hand digraphs respectively. The significant effect of
* digraph class was again confirmed by an analysis of variance, F(3,27) - 11.18, p< .001.

The picture changes somewhat, however, when data from the slow condition is included in the
analyses. The pattern of results across all three conditions is most easily seen by enmining perfor-
mance of individual typists. The overall typing rate and the median interatroke intervals for each
digraph class were calculated separately for each page of text (99 words) in Experiment 4. Figure 3

0
 D. A. Norman and D. E. Rumelhart 79 Final Report, NR 667437

shows these results for two typists who had a wide range of typing rates. The median interstroke inter-
vals are plotted on a logarithmic axis in Figure 3, so that if they maintain a constant relative propor-
tion, the lines should be parallel. The lines are clearly not parallel, illustrating the previous finding
that the data from this experiment do not fit the proportional duration model. Figure 3, however,
shows two patterns that were observed in this study. The first pattern was that at the lowest typing
rate, the median interstroke intervals for all digraph classes were similar, but they diverge as the typing
rate increases. The spread in the median interstroke intervals (as measured by the standard deviation
of the medians) was positively correlated with the typing rate. This was true for all 10 typists on a rela-
tive basis (mean correlation across typists - .96) and for 9 of 10 typists on an absolute basis (mean of
correlation across typists = .70). The second pattern was that at the highest typing rates, the median
interstroke intervals for double and 1-finger digraphs change very little relative to the medians for 2-
finger and 2-hand digraphs. This, of course, is similar to the finding from the rough draft/final copy
study (see Table 6), but now it is clear that the relative amount of change is dependent on the absolute
rate.

Simulaion Model of Typing

Another approach to understanding the control of timing is to construct a simulation model of

performance and investigate the model's predictions of timing. Rumelhart and Norman (1982)
developed a simulation model of a skilled typist that did not have any central timing control. Instead,
the observed timing of the model's keystrokes is the result of competition and cooperation between
efforts to type several letters at once. Similar to skilled typists, the model's interstroke intervals reflect
the constraints of the hands and fingers and the layout of the typewriter keyboard. Although the
model normally attempts to type several letters at once, the letters are normally produced in the
. correct serial order because each letter inhibits every following letter to some extent, and the initial
letter thus usually has the highest activation level. When the level of inhibition between successive
letters is very high, the model essentially types just one letter at a time. As the level of inhibition is
" lowered, however, the model attempts to type several letters at once. In this case, movements to type
successive letters can overlap in time, especially when successive letters are typed by separate fingers or
hands, and the overall typing rate increases. Figure 4 shows how the mean interstroke intervals
decrease with decreasing levels of inhibition. Notice that the different digraph classes do not maintain
constant relative durations, which would be indicated by parallel lines on this logarithmic plot. Dou-
bles do not speed up at all because there is no advantage in moving the hand for the second keystroke;
2,hand and 2-finger digraphs speed up the most because the model can easily overlap movements on
different hands and fingers; and 1-finger digraphs speed up slightly because the model can move its
hand into position for the second keystroke while typing the first. This pattern is similar to the results
from skilled typists for changes near their normal rate (for example, see Table 6) and also matches the
pattern of changes in interstroke intervals observed during acquisition of typing (Gentner, 1983). Note,
however, that the effect of changing the inhibition level in the typing simulation does not completely
match the pattern of results when skilled typists change over wide ranges as in Experiment 4 (se Fig.
ure 3).

. --. *'-.

~~~.........
...-. ,.......................,.. .............
S

D. A. Norman and D. E. Rumelhart 80 Final Report, NR 667-437

Typist t70
500

400
wE

200

300-
oo I , I , I , I * I , I , I
20 30 40 SO O0 V0 0 t0 100

Words per Minute

Typist t74
* 500

400

S300-

200 -

I ! I . I ! I
IL0
to 30 40 G0 s0 S
so o t 00
Words per Minute

Figure 3 Median intestroke intervals for the digraph dmg a a funoctioc of overall typg te. The vrtical a Islog-
arhtbmic. so that itproportional duration modd fit this data, the lines would be pailel. [ncead the totetroke iantervals are
amilar a low rates and quite different at high rates. As typing rate increase the inteutroke intervals for sone dilplb class
contioue to decream. whereas the inteuuroke intervals for double appear to reach a minimum level.

..- . . *.
 -w~-u - . - - 1w Z Ik: K7

D. A. Norman and D. E. Rumelhaul 81 Final Report, NR 667-437

14

I121-
oj g

4
I0
7 0
1.0) .8. 4 2
I Inhbitio Leve

Fiue4 iutsfo b yig iumo ide.Mi netok nevl ortedgqhd a ucino

-. ... .6* *~ *.4 .2*

Inhbiio Leve
D. A. Norman and D. E. Rumelhart 82 Final Report, NR 667437

Multilevel Control of Timing

Models of timing, such as the proportional duration model, seem to be based on the idea that
motor performance is determined solely by central control. With the proportional duration model, a
generalized motor program is stored in long term memory, and with the specification of a multiplica-
tive rate parameter, maps directly onto the timing of behavior, in much the way that the performance
of an ideal robot is a direct reflection of its program. At the other extreme is a model, such as the
typing simulation of Rumelhart and Norman (1982), in which the observed timing is determined pri-
marily by peripheral constraints, such as the structure of the hands. With this model, any central con-
trol is very indirect, for example by changing the inhibition levels between successive letters. Of
course, humans and other creatures can control timing with great precision when the task demands it,
and there is no reason to suspect that people could not use central control to maintain constant rela-
tive intervals if that is required by the task. as in musical performance. But for most action sequences,
the detailed relative timing of components is not a direct part of the task.

Consider the case of a cockroach walking at different rates. The results of Delcomyn (1971) and
Pearson (1972) imply that a motion picture of a fast cockroach, when played back in slow motion,
would not match the performance of a slow cockroach. Why should this be? The only way for cock-
roach to walk slower is by decreasing the angular velocity of its leg during the support phase of the
step cycle when the leg is on the ground. Therefore the movement of the leg is must be directly
related to the speed of walking. In contrast, the leg is not in contact with the ground during the swing
phase, and therefore the angular velocity during the swing phase is not directly constrained by speed of
walking. In fact, unlike the case for the support phase, the angular velocity for the swing phase is rela-
tively constant, perhaps to be close to a minumum energy for the swing phase or maintain stability by
having more feet on the ground when possible. Thus, the ratio of angular velocities for the swing and
support phases (and thus the ratio of durations) changes as the walking speed of the cockroach
changes. In general, when the details of timing are not an explicit part of the task, they may be deter-
mined by other considerations, such as minimizing energy consumption.

The basic point here is that motor performance is an inseparable part of perceptual and cognitive
processes, the physical properties of the body, and the environmental context of the task. The timing
of any particular instance of motor performance is determined by the interaction of thes central and
peripheral processes with environmental task. Results from typing illustrate this multi-level control of
timing. At the central level, we have seen that typists can vary their overall typing rate. Typists also
show effects due to the frequency of words and letter sequences (Larochelle, Gentner, & Grudin, 1985;
Shaffer, 1973; West & Sabban, 1982). At the peripheral level, finger and hand constraints determine
the interstroke intervals of different digraph classes when typing at high speeds. And constraints at
both central and peripheral level interact with the task environment, in this case including issues such
as the particular text to be typed and the layout of the typewriter keyboard. The relative contribution
of all these factors to the final performance depends on the situation. In Experiment 4, when typists
were working at the slowest rates, their timing appears to have been determined by a regular central
timekeeper and all digraph classes were typed at the same rate. As the typists speeded up, however,
peripheral constraints began to play an important role and double and 1-finger digraphs reached their
limiting rate, whereas 2-finger and 2-hand digraphs, where overlapped movement is easier, continued to
met faster.

. ..-.-.
.-..-...... -......".. -.-....- .. ..'..' .
-... -s.'-. -. .- *'- ..-- "--.,
- . . .'- .-.- . -.. - . -. .
D. A. Norman and D. E. Rumelhart 83 Final Report, NR 667-437

Novice typists show a related progression over the course of learning to type. At first their per-
formance appears to be limited by central processes. Instead of following a central timekeeper and typ-
ing all digraph classes at the same rate as a slow expert does, novices appear to be limited by planning
and memory processes. Novices type 1-finger, 2-finger, and 2-hand digraphs at the same rate, but dou-
bles (where there is no need to retrieve a new location for the second key) are typed twice as fast as
the other digraphs (Gentner, 1983). As their skill level increases, novices gradually speed up their typ-
ing, and in the process, the control of timing shifts to the peripheral processes typical of expert typists.

Control of timing is thus determined a several levels in the perceptual-cognitive-motor system,

and the nature and relative importance of these control levels shifts with skill acquisition and the task
environment.
D. A. Norman and D. E. Rumelhart 84 Final Report, NR 667-437

References

Armstrong, T. R. (1970). Trainingfor the production of memorized motor patters (Tech. Rep. No. 26).
Ann Arbor: University of Michigan, Human Performance Center.

Arshavskii, Y. I., Kots, Y. M., Orlovskii, G. N., Rodionov, 1. M., & Shik, M. L. (1965). Investip-
tion of the biomechanics of running by the dog. Bifizika, 10, 665.672.

Carter, M. C., & Shapiro, D. C. (1984). Control of sequential movements: Evidence for generalized
motor programs. Journal of Neurophysiology, 52, 787-796.

Clark, F. J., & Euler, C. von (1972). On the regulation of depth and rate of breathing. Journal of
Physiology, 222, 267-295.

Clarke, E. F. (1982). Timing in the performance of Erik Satic's 'Vexaions'. Acta Psychologica, 50, 1-
19.

Delcomyn, F. (1971). The locomotion of the cockroach, Periplauetaanericana. Journal of Experimental

Biology, 54, 443-452.

Fendrick, P. (1937). Hierarchical skills in typewriting. Journal of Educational Psychology, 28, 609-6M.

Froblich, D. M., & Elliott, J. M. (1984). The schematic representation of effector function underlying
perceptual-motor skill. Journal of Motor Behavior, 16, 40-60.

Gentner, D. R. (1982). Evidence against a central control moici of timing in typing. Journal of Experi-
mental Psychology" Human Perception and Performance, 8, 793-810.

Gentner, D. R. (1983). The acquisition of typewriting skill. Acta Psychologica, 54, 233-248.

Glencross, D. 1. (1973). Temporal organization in a repetitive speed skill. Ergonomics, 16, 765-776.

Goslow, G. E., Reinking, R. M., & Stuart, D. G. (1973). The cat step cycle: Hind limb joint angles
and muscle lengths during unrestrained locomotion. Journal of Morphology, 14), 1-42.

Grillner, S. (1985). Neurobiological bases of rhythmic motor acts in vertebrates. Science, 228, 143-149.

Grillner, S., Halbertsma, J., Nilsson, J., & Thorstensson, A. (1979). The adaptation to speed in
human locomotion. Brian Research, 165, 177-182.

Hollerbach, J. M. (1981). An oscillation theory of handwriting. Biological Cybernetics, 39, 139-156.

Keele, S. W. (1968). Movement control in skilled motor performance. Psychological Bulletin, 70, 387-
403.

4m U

......
 D. A. Norman and D. E. Rumelhart 85 Final Report, NR 667437

Keele, S. W., & Summers, J. J. (1976). The structure of motor programs. In G. E. Stelmach (Ed.),
Motor control: Issues and trends (pp. 109-142). New York: Academic Press.

Klein, R., Levy, S., & McCabe, J. (1984). The parameter preferences of acquired motor programs for
rapid, discrete movements: I. Transfer of training. Memory & Cogrtion, 12, 374-379.

Kozhevnikov, V., & Chistovich, L. (1965). Speech: Articulation and perception. (English translation:
J.P.R.S. Washington, D.C., No. JPRS 30543).

Larochelle, S., Gentner, D. R., & Grudin, J. (1965). Frequency effects In the taracriptionof native ad
foreign language texts. Manuscript in preparation.

Lashley, K. S. (1951). The problem of serial order in behavior. In L. A. Jeffress (Ed.), Cerebral
mechanisms in behavior (pp. 112-136). New York: Wiley.

Lee, D. N., Lishman, J. R., & Thomson, J. A. (8). Regulation of gait in long jumping. Journal of
Experimental Psychology: Human Perceptionand Performance, 448-459.

Macmillan, D. L. (1975). A physiological analysis of walking in the American lobster (homarus ameri-
canus). Philosophical Transactionsof the Royal Society of London, 270, 1-59.

Michon, J. A. (1974). Programs and 'programs" for sequential patterns in motor behavior. Bran
Research, 71, 413-424.

Pearson, K. G. (1972). Central programming and reflex control of walking in the cockroach. Jornal of
Experimental Biology, 56, 173-193.

Prinz, W., & Sanders, A. F. (Eds.) (1984). Cognition and motor processes. Berlin: Springer-Verlag.

Rumelhart, D. E., & Norman, D. A. (1982). Simulating a skilled typist: A study of skilled cognitive-
motor performance. Cognitive Science, 6, 1-36.

Schmidt, R. A. (1975). A schema theory of discrete motor skill learning. PsychologicalReview, 82, 225-
260.

Schmidt, R. A. (1982). Motor control and learning. Champaign, IL: Human Kinetics Publishers.

Schmidt, R. A. (1984). The search for invariance in Ailled motor behavior (Report No. 1111964).
Bielefeld, Federal Republic of Germany. Research Group on Perception & Action at the Centner
for Interdisciplinary Research (ZiP), University of Bielefeld.

Schmidt. R. A., Zelaznik, H. N., Hawkins, B., Frank, J. S., & Quinn, J. T. Jr. (1979). Motor-
output variability: A theory for the accuracy of rapid motor acts. Psykhclogicd Review, N, 415-451.

Shaffer, L. H. (1973). Latency mechanisms in transcription. In S. Koreblum (Ed.), Anention d

%
, • -. -. .. ° '° - . -- " . --
". . ° .. . . .• . . . . .-.- - .oq " . " . ~
 D. A. Norman and D. E. Rumelhart 86 Final Report, NR 667437

performance IV (pp. 435-446). New York: Academic Press.

Shaffer, L. H. (1984). Timing in solo and duet piano performances. The Quarterly Journal of Experi-
mental Psychology, 36A, 577-595.

Shapiro, D. C. (1976). A preliminary attempt to determine the duration of a motor program. In D.

M. Landers & R. W. Christina (Eds.), Psychology of sport and moor behavior I (pp. 17-24). Cham-
paign, IL: Human Kinetics.

Shapiro, D. C., & Schmidt, R. A. (1982). The schema theory: recent evidence and developmental
implications. In J. A. S. Kelso & J. E. Clark (Eds.), The development of movement control and coor-
dination (pp. 113-150). New York: John Wiley.

Shapiro, D. C., Zernicke, R. F., Gregor, R. J., & Diestel, J. D. (1981). Evidence for generalized
motor programs using gait pattern analysis. Journalof Motor Behavior, 13, 33-47.

Sternberg, S., Knoll, R. L., Monsell, S., & Wright, C. E. (1983). Control of rapid action sequences in
speech and typing. Paper presented at the annual meeting of the American Psychological Associa-
tion.

Sternberg, S., Monscil, S., Knoll, R. L., & Wright, C. E. (1978). The latency and duration of rapid
movement sequences: Comparisons of speech and typewriting. In G. E. Stelmach (Ed.), liforma-
lion processing in motor control and learning(pp. 117-152). New York: Academic Press.

Summers, J. 1. (1975). The role of timing in motor program representation. Journal of Motor Behavior,
7,229-241.

Terzuolo, C. A., & Viviani, P. (1979). The central representation of learned motor patterns. In R. E.
Talbott & D. R. Humphrey (Eds.), Posture and movement (pp. 113-121). New York: Raven Press.

Tcrzuolo, C. A., & Viviani, P. (1980). Determinants and characteristics of motor patterns used for
typing. Neuroscience, 5, 1085-1103.

Tullcr, B., Harris, K. S., & Kelso, J. A. (1981). Articulatory motor events as a function of speaking rae
and stress (Haskins Laboratories: Status Report on Speech Research SR-65).

Tuller, B., Kelso, J. A. S., & Harris, K. S. (1981). Phase relntionships among articulator muscles as a
function of speaking rare and stress (Haskins Laboratories: Status Report on Speech Research SR-65).

Tuller, B., Kelso, J. A. S., & Harris, K. S. (1982). Inte'articulator phasing as an index of temporal
regularity in speech. Journal of Experimental Psychology: Human Perception and Performance, 8, 460.
472.

Viviani, P., & Terzuolo, C. (1980). Space-time invariance in learned motor skills. In G. E. Stelmach
& J. Requin (Eds.), Tutorials in motor behavior (pp. 525.533).

..1P . ., . . . - .. o , ' " . . , -- " . " . ., ..-. . -. ' . . -. - , ' . .. - . - .- . . -.-. -..., . . -. - . . - . .
D. A. Norman and D. E. Rumelhart 87 Final Report, NR 667-437

West, L. I., & Sabban, Y. (1982). Hierarchy of stroking habits at the typewriter. Journalof Applied
Psychology, 67, 370-376.

Williams, T. L. (1977). Tortoise locomotion. Journal of Physioloy (London), 270, 54-55.

Wing, A. M. (1978). Response timing in handwriting. In G. E. Stelmach (Ed.), Iformation processing

in motor control and learning (pp. 153-172). New York: Academic Press.

Wing, A. M. (1980). The long and short of timing in response sequences. In G. E. Stelmach & J.
Requin (Eds.), Titorials in motor behavior (pp. 469-486).

Winter, D. A. (19&1). Biomechanical motor patterns in normal walking. Journal of Motor Behavior, 15,
302-330.

Zelaznik, H. N., Schmidt, R. A., Gielen, S. C. A. M., & Milich, M. (1985). Kinematic properties of
rapid aimed hand movements. Manuscript submitted for publication.

r .... ...i" '".

" '- ' , ' " '. , '" ," . ..-.. .- . . . . . . . . ".. . '.'-'-
. . ..--.
'. , .,-" ... L ,-- -G
D. A. Norman and D. E. Rumelhart 88 Final Report, NR 667437

PUBLICATIONS
Technical Reports

ONR-80Nl. Donald R. Gentner, Jonathan Grudin, and Eileen Conway. Finger Movements in
TranscriptionTyping. May 1980.

ONR-8002. James L. McClelland and David E. Rumelhart. An Interactive Activation Model of the
Effect of Context in Perception: Part I. May 1980. Also published in Psychological
Review. 88.5, pp. 375-401, 1981.

ONR-80W3. David E. Rumelhart and James L. McClelland. An Interactive Activation Model of the
Effect of Context in Perception: Part II. July 1980. Also published in Psychological
Review. 89, 1, pp. 60-94, 1982.

ONR-8004. Donald A. Norman. Errorsin Human Performance. August 1980.

ONR-8005. David E. Rumelhart and Donald A. Norman. Analogical Processes in Learning. Sep-
tember 1980. Also published in J. R. Anderson (Ed.), Cognitive skills and their acquisi-
tion. Hillsdale, NJ: Erlbaum, 1981.

ONR-8006. Donald A. Norman and Tim Shallice. Attention to Action: Willed and Automatic Con-
trol of Behavior. December 1980.

ONR-8101. David E. Rumelhart. Understanding Understanding. January 1981.

ONR-8102. David E. Rumelhart and Donald A. Norman. Simulating a Skilled Typist: A Study of
Skilled Cognitive-Motor Performance. May 1981. Also published in Cognitive Science, 6,
pp. 1-36, 1982.

ONR-8103. Donald R. Gentner. Skilled Finger Movements in Typing. July 1981.

ONR-8104. Michael I. Jordan. The Timing of Endpoints in Movement. November 191.

ONR-8105. Gary Pelman. Two Papers in Cognitive Engineering: The Design of an Interface to a
ProgrammingSystem and MENUNIX: A Me-&uzed Interface to UNIX (User Mald).
November 1961. Also published in Proceedings of the 1982 USENIX Coolerence, San
Diego, CA, 1982.

...................................... ,....... . o . - . a
D. A. Norman and D. E. Rumelhart 89 Final Report, NR 667437

ONR-8106. Donald A. Norman and Diane Fisher. Why Alphabetic Keyboards Are Not Easy to Use:
Keyboard Layout Doesn't Much Matter. November 1981. Also published in Human Fac-
tors, 24, pp. 509-515, 1982.

ONR-8107. Donald R. Gentner. Evidence Against a Central Control Model of Timing in Typing.
December 1981. Also published in Journal of Experimental Psychology: Human Percep-
tion and Performance,8, pp. 793-810, 1962.

ONR-8201. Jonathan T. Grudin and Serge Larochelle. Digraph Frequency Effects inSkilled Typ-
ing. February 1982.

ONR-82f2. Jonathan T. Grudin. Central Control of Timing 4n Skilled Typing. February 1962.

ONR-8203. Amy Geoffroy and Donald A. Norman. Ease of Tapping the Fingers in a Sequence
Depends on the Mental Encoding. March 1982.

ONR-8204. LNR Research Group. Studies of Typing fron the LNR Research Group: The role of con-
text, differences in skill level, errors, hamd movements, and a computer simulation. May
1982. Also published in W. E. Cooper (Ed.), Cognitive aspects of skilled typewriting.
New York: Springer-Verlag, 1983.

ONR-82W5. Donald A. Norman. Five Papers on Human-Machine Interaction. May 162. Also pub-
lished individually as follows: Some observations on mental models, in D. Gentner
and A. Stevens (Eds.), Mental models, Hillsdale, NJ: Erlbaum, 1983; A psychologist
views human processing: Human errors and other phenomena suggest processing
mechanisms, in Proceedings of the Internaional Joint Conference on Artificial Intelli-
gence, Vancouver, 1981; Steps toward a cognitive engineering- Design rules based on
analyses of human error, in Proceedings of the Conference on Human Factors in Com-
puter Systems, Gaithersburg, MD, 1982; The trouble with UNIX, in Daramaion, 27.12,
November 1981, pp. 139-150; The trouble with networks, in Datonation, January 1982,
pp. 188-192.

ONR-82M. Naomi Miyake. Constructive Interaction. June 1982.

ONR-8207. Donald R. Gentner. The Development of Typewriting Skill. September 1962. Also
published as Acquisition of typewriting skill, in Acta Psychologica,54, pp. 233-248, 1983.

ONR-82M8. Gary Perlman. Natural Artificial Languages: Low-Level Processes. December 1982.
Also published in The International Journal of Man-Machine Studies, 20, pp. 373-419,
1984.

ONR-8301. Michael C. Mozer. Letter Migration in Word Perception. April 1983. Also published in
Journal of Experimental Psycholog: Humnan Perceptionand Performace.9,4, pp. 531-546,
1983.

....................... -.... , ....

D. A. Norman and D. E. Rumelhart 90 Final Report, NR 667-437

ONR-8302. David E. Rumelhart and Donald A. Norman. Representaion in Memory. June 1983.
To appear in R. C. Atkinson, G. Lindzey, & R. D. Luce (Eds.), Handbook of experi-
menial psychology. New York: Wiley (in press).

ONR-8303. The HMI Project at University of California, San Diego. User Centered System Design:
Part 1, Papers for the CHI 1983 Coference an Human Factors in Computer Systems.
November 1983. Also published in A. Janda (Ed.), Proceedings of the CHI '83 Confer-
ence on Human Factors in Computing Systems. New York: ACM, 1983.

ONR-8304. Paul Smolensky. Harmony Theory: A Mathematical Framework for Stochastic Parallel
Processing. December 1983. Also published in Proceedingsof the National Conference
on Artificial Intelligence, AAAI.83, Washington DC, 1983.

ONR-8401. Stephen W. Draper and Donald A. Norman. Software Engineeringfor User Interfaces.
January 1984. Also published in Proceedingsof the Seventh InternaionalConference on
Software Engineering, Orlando, FL, 1984.

ONR-8402. The UCSD HMI Project. User Centered System Design: Part II, Collected Papers.
March 1984. Also published individually as follows: Norman, D.A. (1984), Stages and
levels in human-machine interaction, International Journal of Man-Machine Studies, 21,
365-375; Draper, S.W., The nature of expertise in UNIX; Owen, D., Users in the real
world; O'Malley, C., Draper, S.W., & Riley, M., Constructive interaction: A method
for studying user-computer-user interaction; Smolensky, P., Monty, M.L., & Conway,
E., Formalizing task descriptions for command specification and documentation;
Bannon, L.J., & O'Malley, C., Problems in evaluation of human-computer interfaces:
A case study; Riley, M., & O'Malley, C., Planning nets: A framework for analyzing
user-computer interactions; all published in B. Shackel (Ed.), INTERACT '84, First
Conference on Human.Compuaer Interaction, Amsterdam: North-Holland, 1984; Norman,
D.A., & Draper, S.W., Software engineering for user interfaces, Proceedings of the
Seventh InternationalConference on Software Engineering,Orlando, FL, 1984.

ONR-8403. Paul Smolcnsky and Mary S. Riley. Harmony Theory: Problem Solving. Parallel Cogni-
tive Models, and Thermal Physics. April 1984. The first two papers appear in Proceed-
ings of the Sixth Annual Meeting of the Cognitive Science Society, Boulder, CU, 1984.

ONR-8404. Donald R. Gentner. Expertise in Typewriting. April 1984. To appear in M. T. Chi

(Ed.), The nature of expertise (in press).

ONR-8405. Michael C. Mozer. Inductive Informaion Retrieval Using Parallel Distributed Comipua-
tion. May 1984.

ONR-8M06. David E. Rumelhart and David Zipser. Featre Discovery by Compedive Learming.
July 1984. Also published in Cognitive Science, 1985, 9, 75-112.

%( -' .-. - - , .. . .. . . -.- - . . .

kIDIEGO
5
AD-fl157 66

UNCLASSIFIED
HUMAN COGNITION AND PERFOR ANCE(U) CALIFORNIA UNIV SAN
LA JOLLA INST FOR COGNITIVE SCIENCE
D A NORMAN ET AL. 01 MAY 85 N010014-79-C-032.3
F/G 5/10
2/2

NL
 .5

-a.a

13

.4 - .-
I
1111 - In2 III,25 Sl

-1111111

1111
~MICROCOPY RESOLUTION TEST CHART
1-111 [8
.:,.jNATIONALBUREAU
STANDARDS-]963-A
OF
L 136%
mu 1.%

%- h

'.a-;-,

'F.'.;

:.,

h.r ', ' ; . ' ,,' 3 - ":'. . .:: , -- ".',.. . . ," .".... .: . . . ; .,.. . ...:.:., , ; ,..
 D. A. Norman and D. E. Rumelhart 91 Final Report, NR 667-437

Archival Publcations

Miyake, N., & Norman, D. A. (1979). To ask a question, one must know enough to know what is not
known. Journal of Verbal Learning and Verbal Behavior, 18.357-364.

Norman, D. A. (1979). Copycat science. In R. A. Cole (Ed.), Perception ad production of fluent

speech. Hillsdale, NJ: Erlbaum.
Norman, D. A. (1979). Perception, memory, and mental processes. In L. 0. Nilsson (Ed.), Perspec-

lives on memory research. Hillsdale,NJ: Erlbaum.

Norman, D. A., & Bobrow, D. G. (1979). Descriptions: An intermediate stage in memory retrieval.
Cognitive Psychology, 11, 107-123.

-,. "Norman, D. A. (1980). Cognitive engineering and education. In D. T. Tuma & F. Reif (Eds.), Problem
solving and education:Issues in teaching ad research. Hillsdale, NJ: Erlbaum.

Norman, D. A. (1980). Teaching, learning, and the representation of knowledge. In R. E. Snow, P. A.

Federico, & W. E. Montague (Eds.), Aptitude, learning, and intruction: Vol. 2. Cognitive process an-
lyses of learningand problem solving. Hillsdale, NJ: Erlbaum.

Norman, D. A. (1980). Twelve issues for cognitive science. Cognitive Science, 4. 1-32.

Norman, D. A. (1980). What goes on in the mind of the learner? In W. . McKcachie (Ed.), Cognition,
.*-. college teaching, and student learning. San Francisco: Jossey-Bass.

Norman, D. A. (1980). Why people make mistakes. Reader's Digest, 117. 103-106.

Perman, G. (1980). Data analysis programs for the UNIX operating system. Behavior Research Methods
and Instrumentation,12, 554-558.

Rumeihart, D. E. (1980). On evaluating story grammars. Cognitive Science, 4,313-316.

Rumelhart, D. E. (1980). Schemata: The building blocks of cognition. In R. Spiro, B. Bruce, & W.
*L Brewer (Eds.), Theoretical issues In reading comprehension. Hillsdale, NJ: Erlbaum.

Erickson, T. D., & Mattson, M. E. (1981). From words to meaning: A semantic illusion. Jownal of
*
"' Verbal Learning and Verbal Behavior. 20, 540-551.

McClellhnd, J. L., & Rumelhart, D. E. (1961). An interactiv activation model of the effect of contet
in perception: Part 1. An account of basic findinp. Psychologlca RevIew, 88. 375407.

Norman, D. A. (1981). Categorization of action dips. Psychologica Revew, 8.,1-15.

Norman, D. A. (Ed.). (1981). Perspectives on cognitive xience: The La Jolie Cufereme. Published
 D. A. Norman and D. E. Rumelhart 92 Final Report, NR 667437

jointly by Eribaum and Ablex. Norwood, NJ: Ablex. Hillsdule, NJ: Eribaum.

Norman, D. A. (1981). A psychologist views human processing- Human errors and other phenomena
suggest processing mechanisms. Proceedings of! uhe InteraioncalJoint Coiference on Artificial Intelli-
gence, Vancouver.

Norman, D. A. (1981). The trouble with UNIX: The user interface is horrid. Daransaion,27, 238-15.

Norman, D. A., & Rumeihart, D. E. (1981). The LNR approach to human information processing.
Cognition, 10, 235-240.

Rumeihart, D. E. (1981). Analogical processes in learning. In J. R. Anderson (Ed.), Cognitive Skills

and their acquisition. Hillsdale, NJ: Eribaum.

Rumclhart, D. E., & Norman, D. A. (1981). An activation-trigger-schema model for the simluation of
skilled typing. Proceedings of the Berkeley Cost erence of the Cognitive Science Society.

0 Rumelhart, D. E., & Norman, D. A. (1981). Parallel associative models of memory: Where do they
fit? In G. Hinton & J. Anderson (Eds.), Parallelassocitive models of memory. Hillsdale, NJ: ErI-
baum.

Gentner, D. R. (1982). Acquisition of typewriting skill. Adta Psyclaologia.54. 233-248.

Gentner, D.R. (1982). Evidence against a central control model of timing in typing. Joura of Experi-
mental Psychology: Human Perception and Performance, 8. 793-810.

Gentner, D. R. (1982). Testing the central control model of typing: Comments on the reply by Viviani
and Terzuolo. Journal of Experimental Psychology: Humnan Perception and Performance, 8,814-816.

Miyake, N. (1982). Mental models. Psychology, 3 (3), 12-19. (in Japanese.)

Norman, D. A. (1982). Comments on cognitive engineering, the need for clear wsystem images,' and the
safety of nuclear power plants. Proceedings of IEEE Standards Workshop on Humsan Factors ad
Nuclear Safety. New York: IEEE.

Norman, D. A. (1982). How to make computer systems that people will like to use: Steps toward a
cognitive engineering. Proceedingsof the Western Electronic Show and Convention (WescoJU2).

Norman, D.A. (1982). Learning and memory. San Francisco: Freeman. (Translations: Japanese, in pro-
* gross.)

Norman, D. A. (1982). Steps towards an cognitive engineering: Design rules based on analyses of
human error. Proceedings of the Ceqference on Human Factors In Computer Systems. Gait hersburg,
MD.
plp'"Np~~~~C --- ---- W ---- -- -- '-rr.. '- - . - - ------ IF -W- t--V V--

D. A. Norman and D. E. Rumeihart 93 Final Report, NR 667-437

Norman, D. A. (1982). The trouble with networks. Datamatio, zz, 188-192.

Perlman, 0. (1982). Data analysis in the UNIX environment: Automated experimental design specifi-
cation for data analysis programs. Proceedings of the XIV CoVerence an the Interface, Troy, NY.

Perlman, Gary, (1982). Experimental control programs for the UNIX operating system. Behavior
Research Methods and Instrumentation,14, 417-421.

Rumelhart, D. E., & McClelland, J. L. (1982). An interactive activation model of the effect of context
in perception: Part 2. The contextual enhancement effect and some test and extensions of the
model. Psychological Review, 89, 60-94.

Gentner, D.R. (1983). Keystroke timing in transcription typing. In W. E. Cooper (Ed.), Cogntive
aspects of skilled typewriting. New York: Springer-Verlaag.

Gentner, D. R., Grudin, J., Larochelle, S., Norman, D. A., & Rumelhart, D. E. (1963). A glossary of
terms, including a classification of typing errors. In W. E. Cooper (Ed.), Cognitive aspects of skilled
typewriting. New York: Springer-Verlag.

Grudin, Jonathan T. (1983). Error patterns in skilled and novice transcription typing. In W. E.
Cooper (Ed.), Cognitive aspects of skilled typewriting. New York: Springer-Verlag.

Larochelle, Serge (1983). A comparison of skilled and novice performance in discontinuous typing. In
W.E. Cooper (Ed.), Cognitive aspects of skilled typewriting. New York: Springer-Verlag.

Mozer, Michael C. (1983). Letter migration in word perception. Jowmal of Experimental Psychology:
Human Perception and Performance,9, 531-546.

Norman, D. A. (1983). Design principles for human-computer interfaces. In A. Janda (Ed.), Proceed-
ings of the CHI '83 Coference on Human Factorsin Conputing Systems. New York: ACM.

Norman, D. A. (1983). Design rules based on analyses of human error. Commwdcwlow of the ACM, 4.
254-258.
(Also translated into Japanese and printed in Psychology (1983), 41, No. 8, 62-69.)
Originally printed in Proceedings of the Coference an Human Factors in Compter Systems. (March,
1982). Gaithersberg, MD: Institute for Computer Sciences and Technology, National Bureau of
Standards, U.S. Department of Commerce.

Norman, D.A. (1963). Some observations on mental models. In D. Gentner & A. Stevens (Eds.),
Mental Models. Hillsdale, NJ: Erlbaunm.

Norman, D.A. (1983). Theories and models in cognitive psychology. In E. Douclkin (Ed.), Conitive
psychohysiology. Hillsdale, NJ: Erlbaum.

Norman, D. A. (1983). Worsening the knowledge gap: The mystique of computation builds
D. A. Norman and D. E. Rumelhart 94 Final Report, NX 667437

unnecessary barriers. Annals of the New York Academy of Sciences, 246, 22D-233.
Norman, D. A., (1983, September). The DVORAK revival: Is it really worth the cost? cpnews: Comsu-
mer Products Tech Group: the Human FactorsSociety. 8, No. 3, 5-7.

Norman, D. A. (1983, October). The computer always rinp twice. P ycholojyToday. pp. 46-50.

Norman, D. A., & Rumethart, D. E. (1983). Studies of typing from the LNR Research Group. In W.
E. Cooper (Ed.), Cognitive aspects of skilled typewritiq. New York: Springer-Verlag.

Rumelhart, D. E., & McClelland, J. L. (1983). Interactive processing through spreading activation. In
C. Perfetti & A. Lesgold (Eds.), Interactive processes in reading. Hillsdale, NJ: Eribaum.

Compton, M. (1984, November). Don Norman speaks out: An interview with the UNIX system's best
known critic. UNIX Review, pp. 38-49, 94-96.

Draper, S.W., & Norman, D. A. (1984). Software engineering for user interfaces. Proceedings of the
7th InternationalConference on Software Engineering. Orlando, FL.

Gentner, D. R., & Norman, D. A. (1984, March). The typist's touch. Psychology Today, pp. 66-72.

Norman, D. A. (1984). Cognitive engineering principles in the design of human-computer interfaces.

In G. Salvendy (Ed.), Hunan-Computer Interaction:Proceedings of the First USA-Jqpon Cooference
on Human-Computer Interaction (Honolulu,August 1984). New York: Elsevier.

Norman, D. A. (1984). Stages and levels in human-machine interaction. InternationalJowral of Man.

Machine Studies, 21, 365-375.

Perlman, Gary, (1984). Natural artificial languages: Low-level processes. International Journal of Man-
Machine Studies, 20, 373-419.

Riley, M. S., & Smolensky, P. (1984). A parallel model of (sequential) problem solving. Proceedingsof
the Sixth Annual Conference of the Cognitive Science Society. Boulder, CO.

Rumelhart, D. E. (1984). Understanding underistanding. In J. Flood (Ed.), Understoding reading

comprehension. Newark, DE: International Reading Asociation.

Smolensky, P. (1984). The mathematical role of self-consistency in parallel computation. Proceedinp

of the Sixth Anmal Corderenceof the Cognitive Science Society. Boulder, CO.

Bannon, L. J., & O'Malley, C. E. (1985). Problems in evaluation of human-computer interfaces: A

cae study. In B. Shackel (Ed.), INTERACT '84: First Ctference on HwnmamCmputer Interkdan.
Amsterdam: North-Holland.

Draper, S. W. (1985). The nature of eapertise in UNIX. In B. Shackel (Ed.), INTERACT '84: First

.........
 D. A. Norman and D. E. Rumelhart 95 Final Report, NR 667437

Conference on Human.Computer Interaction. Amsterdam: North-Holland.

Norman, D. A. (1985). Stages and levels in human-machine interaction. In B. Shackel (Ed.),

INTERACT '84: First Conference on Hmwan-Comuer Interaction. Amsterdam: North-Holland.

O'Malley, C. E., Draper, S. W., & Riley, M. S. (1985). Constructive interaction: A method for study-
ing user-computer-user interaction. In B. Shackel (Ed.), INTERACT '8f: First Coference on
Human-Computer Interaction. Amsterdam: North-Holland.

Owen, D. (1985). Users in the real world. In B. Shackel (Ed.), INTERACT '8f: First Conference on
Human.Computer Interaction. Amsterdam: North-Holland.

Riley, M. S., & O'Malley, C. E. (1985). Planning nets: A framework for analyzing user-computer
-:.' interactions. In B. Shackel (Ed.), INTERACT '84: First Coference n HumanCompurerInteraction.
Amsterdam: North-Holland.

Rumelhart, D. E., & Zipser, D. (1985). Feature discovery by competitive learning. Cognitive Science, 9,
75-112.

Smolensky, P. Monty M. L., & Conway, E. (1985). Formalizing task descriptions for command specifi-
cation and documentation. In B. Shackel (Ed.), INTERACT '84: First Conference on Human-
Computer Interaction. Amsterdam: North-Holland.

McClelland, J.L., & Rumelhart, D. E. (in press). Parallel distributed processing: Explorations in the
microstructure of cognition. Vol. 2: Applications. Cambridge, MA: MIT Press/Bradford Books.

V.,, Norman, D. A., & Draper, S. W. (in press). User center system design: New perspectives In human-
computer interaction. Hillsdale, NJ: Erlbaum.

Rumelhart, D. E., & McClelland, J.L. (in press) Parallel distributed processing: Explorations in the
microstructureof cognition. Vol. 1: Foundations. Cambridge, MA: MIT Press/Bradford Books.

Rumelhart, D.E., & Norman, D.A. (in press). Representation in memory. In R.C. Atkinson, L.J.
Herrnstein, G. Lindzey, & R.D. Luce (Eds.), Handbook of Experimental Psychology. NY: Wiley.
TOL
 ONR Distributim Lst

E. Aiken James Lester

Navy Personnel Research Office of Naval Research
& Development Center 800 North Quincy Street
San Diego, CA 92152 Arlington, VA 22217

Robert Bachman Harold F. O'Neil, Jr.

ONR, 0-043 Army Research Institute
University of California, San Diego 500 Eisenhower Avenue
La Jolla, CA 92D93 Alexandria, VA 22333

Susan Chipman Joe Psotka

Personnel & Training Research (Code 442PT) Army Research Institute
Office of Naval Research 5001 Eisenhower Avenue
800 North Quincy Street Alexandria, VA 22333
Arlington, VA 22217
Michael Shafto
Charles E. Davis Personnel & Training Research (Code 442PT)
Personnel & Training Research (Code 442PT) Office of Naval Research
Office of Naval Research 800 North Quincy Street
800 North Quincy Street Arlington, VA 22217
Arlington, VA 22217
John Tanguey
Craig Fields Program Manager
Advanced Research Life Sciences Directorage
Projects Agency Bldg 410
1400 Wilson Blvd. AFOSR/NL
Arlington, VA 22209 Boiling Air Force Base
Washington, DC 20332
Eugene Gloye
ONR Branch Office Jack Thorpe
1030 E. Green St. Advanced Research Projects
Pasadena, CA 91101 Agency
1400 Wilson Blvd.
James Hollan Arlington, VA 22209
Code 13
NPRDC Wallace Wulfck
San Diego, CA 92152 Code 13
NPRDC
San Diego, CA 92152

....... ............ ................ "

F% T-s--.-
-v-r--. i

FILMED
9-85

DTIC -V i