|
2005 > |
Course
Description |
Introduction
to Cognitive Modelling and
Cognitive Architectures
Richard M Young
UCLIC: University
College Lo ndon Interaction Centre
Computational cognitive modelling lies at the heart of
cognitive science. It involves the
construction of computer programs which embody theory and explain the cognitive
processes underlying everyday and laboratory behaviour.
The course will deal with symbolic and hybrid approaches to
cognitive modelling, though not connectionist (or neural net) approaches. We will discuss the notion of cognitive architectures and examine
their importance for explaining cognition and constructing unified
theories. In the practical sessions, we
will become familiar with the COGENT language and its use for building
cognitive models.
Textbooks
The field of cognitive modelling is not well served by
general textbooks, especially for the symbolic approach. If you pursue the subject further, I
recommend that you consider looking at the following:
Cooper, R. P. (2002) Modelling
High-Level Cognitive Processes.
Lawrence Erlbaum Associates.
This is really the first textbook on
cognitive modelling to exist. It
describes and uses the COGENT modelling language. Chapter 1 is one of the best general
introductions to cognitive modelling available.
Sec 1.9, on different approaches to cognitive modelling, is especially
good. Chapter 9 provides a good, critical
look at the state of the art in cognitive modelling.
Anderson,
J. R. & Lebiere, C. (1998) The Atomic
Components of Thought. Erlbaum.
See comment below under Lecture 4.
Newell,
A. (1990). Unified Theories of Cognition. Harvard University Press.
See comment below under Lecture 3.
Polk, T. A. & Seifert, C. M. (2002) Cognitive
Modeling. MIT Press.
This is not a textbook, but a useful collection of key
readings.
For an introduction to connectionist modelling (which we
will not cover in this course), I recommend the following, though there are
several other good textbooks available too.
You might like to see also my review
of the book (Young, R. M. (2003).
Review of P. McLeod, K. Plunkett & E. T. Rolls, Introduction
to Connectionist Modelling of Cognitive Processes. Psychology
Learning and Teaching, 3
(1), 57-58.)
Lecture 1. Computer models and why
they are needed
* brief historical background
* characteristics of cognitive models
* symbolic vs connectionist architectures
Required readings:
This is a much-quoted paper, and is rightly
regarded as the manifesto for the whole idea of devloping cognitive
architectures and trying to use them as unified theories.
Optional readings:
Newell, A. (1990). Introduction.
Chapter 1 of A. Newell, Unified
Theories of Cognition, pp.1-41.
Harvard University Press. (hardcopy)
Introduces what happens, some 15-20 years later,
if one pursues the ideas in the "20 questions" paper. Read it all.
Even where it diverges a bit from our path, it's good for your
soul. (This is also the supplementary
reading for Day 3.)
Cooper, R. P. (2002) Modelling
High-Level Cognitive Processes, Chapter 1.
Lawrence Erlbaum Associates.
See comment above under Textbooks.
Lecture 2. Production systems
* production system architectures
* characteristics of production systems
* example models
Required readings:
A short encyclopedia entry, with some key references.
Young, R. M. & O’Shea, T.
(1981) Errors in children’s
subtraction. Cognitive Science, 5, 153-177.
There are notes available to guide
you through the reading.
Optional readings:
This introduction is rather dated now,
but it does explain clearly the basic concepts of production systems. There are
notes available to guide you through the reading.
Lecture 3. Cognitive architectures
* the argument for unified theories
* ACT-R’s “no magic” properties
* relation between a cognitive model and
its cognitive architecture
Required readings:
Anderson, J. R. & Lebiere,
C. (1998) The No-Magic doctrine in
ACT-R. The Atomic Components of Thought, pp.14-17.
Erlbaum. (hardcopy)
This short excerpt from the ACT-R book
lays out the “no magic” principles which we will be going through in the
lecture.
This very short paper (1000 words) is a
commentary on a target article by John Anderson& Christian Lebiere. Don't worry about the first paragraph or so,
but the point of recommending it to you is because it summarises very briefly
the argument in Howes & Young (1997), the paper mentioned immediately
below. You'll have to judge for
yourselves whether the longer or the shorter version is the clearer.
The main aim of the paper is to explore
the extent to which an architecture has its own theoretical content, which
constrains any models built within it.
The architecture used is Soar, but the paper is intended to be
self-contained. If you can't read it
all, skip sections 2.2.1-2 and 3.3-6.
Optional readings:
Newell, A.
(1990). Introduction. Chapter 1 of A. Newell, Unified Theories of Cognition, pp.1-41. Harvard University Press.
(Chapter 1 is the supplementary
reading for Day 1: please see comments above.)
The book as a whole is now a bit dated, mainly because Soar, Newell’s
choice of candidate unified theory, has been overtaken by ACT-R. But it remains a superb book, and chapters 1,
2, 3, and 8 cover a lot of material relevant to the course.
Lecture 4. Introduction to ACT-R
* system structure
* declarative information
* procedural information
* example model
Required readings:
This recent paper summarises the scope and power of ACT-R.
Optional readings:
Anderson, J. R. & Lebiere,
C. (1998) The Atomic Components of
Thought. Erlbaum.
This is an excellent book, and is the
best and clearest introduction to ACT-R.
Unfortunately, it cannot be treated as a textbook because it describes version
4 of ACT-R, whereas the current version is now up to version 6. Nonetheless, it is worth reading and
describes some of the best research done with ACT-R up to the date of
publication. For our short course, the
Preface, Chapter 1, and part of Chapter 2 (pp.19-29) are relevant. Chapter 12 provides a useful summary and
overall perspective.
I've included this paper because it's a
first-rate illustration of what can be achieved with ACT-R. The authors model a task, and fit the
performance data of individual Ss by setting a single parameter, W = source
activation, for each S. Then, using the
same individual value of W for each S, they predict a priori Ss'
performance on a second task. Really
excellent cognitive science.
http://act-r.psy.cmu.edu/tutorials/
The ACT-R website includes an excellent
set of tutorial material for learning the language, including tutorial units,
exercises, and example models. As well
as the tutorials, there are lists of publications and published models using
ACT-R.
Lecture 5 Methodological
aspects
* verbal protocols
* averaging across strategies
* evaluating cognitive models
Required readings:
The paper provides a strong and clear example of how
averaging over different strategies can yield misleading results.
Optional readings:
Ericsson, K. A. & Simon,
H. A. (1980). Verbal reports as
data. Psychological Review, 87, 215-251.
This paper discusses in what ways and
under what conditions talk-aloud protocols can validly be treated as
psychological data. It provides a simple
framework for thinking about verbal reports, thereby enabling the discussion
and recommendations to be set in the context of a theory of the cognitive
processes by which a person generates the verbal protocol. Much of the paper is taken up with a discussion
of the empirical evidence of the effects of verbalisation on cognitive
processing, which is not immediately relevant for our purpose. I suggest reading at least the first part of
the paper, through to p.226 (and include the RH column too) and then the General
Discussion on p.247. For a fuller
treatment, see the book (listed next), but the paper provides a good summary.
See comment on the reading above.
van Someren, M. W., Barnard,
Y. F. & Sandberg, J. A. C. (1994). The Think Aloud Method: A Practical Guide to
Modelling Cognitive Processes.
Academic Press.
One of the few books explicitly about
cognitive modelling, or at least a specialised aspect thereof. The book claims to be relevant to modelling
for both Knowledge Engineering (i.e. building expert systems) and psychological
investigation, though to my eye it tilts more towards the Kn Engineering. It's good on practicalities and for a summary
discussion of applicability of think-aloud technique, and can perhaps be
consulted as an alternative to Ericsson & Simon. But the short section on validation
(pp.131-134) is almost useless.
Assignments
Students
who take the course for credit will be assigned a practical exercise in using
the COGENT cognitive modelling language that will be taught in the practical
sessions.
Richard M. Young
Richard M.
Young received his PhD from Carnegie Mellon University in 1973. He worked at the Department of Artificial
Intelligence at Edinburgh University, and then in 1978 moved to the Medical
Research Council’s Applied Psychology Unit in Cambridge, UK. From 1997 he was Professor of Cognitive
Science at the University of Hertfordshire (UK), and he recently moved to
UCLIC, the University College London Interaction Centre. He is Chair of the Cognitive Science Society
2005-6. His research interests centre on
the computer modelling of human cognition, and include the topics of mental
models, the learning and performance of complex tasks, and people’s exploratory
use of interactive devices.