Introduction to the ACT-R cognitive architecture

Dieter P Wallach <dwallach+@andrew.cmu.edu>

8 Dec 1997


This tutorial will cover the fundamentals of the ACT-R architecture. 
ACT-R is a cognitive architecture based on experimental results and
theoretical assumptions about the acquisition, organization and
application of human knowledge.  It is a fully implemented simulation
system that has been applied to tasks ranging from simple reaction time
to complex air traffic control.  ACT-R distinguishes between a symbolic
and a subsymbolic level.  The symbolic level rests on the idea of a
production system that differentiates between permanent representations
of declarative and procedural knowledge.  Declarative knowledge is
represented in the form of "chunks" while procedural knowledge is
encoded in productions.  Underlying the symbolic level, a subsymbolic
level based on the concept of "Rational Analysis" (Anderson, 1990)
allows the architecture to adapt to the statistical structure of the
environment.  The tutorial describes the most recent version of ACT-R as
it will be published in the forthcoming book "The Atomic Components of
Thought" by J.R. Anderson & C. Lebiere.  The most relevant text for the
tutorial is the book "Rules of the mind" by J.R. Anderson (Erlbaum,
1993). A short introduction to ACT-R is provided in "ACT-R: A simple
theory of complex cognition" (Anderson, 1995: American psychologist, 51
(4), pp. 355-365).

The tutorial will be organized around the following sections, each of
which will be illustrated by example ACT-R models:

(1) Introduction
This section introduces the basic concepts of ACT-R and provides an
overview of the overall system.  The structure of chunks and productions
is described and the pattern matching process that relates declarative
and procedural knowledge is outlined.  Finally the concept of goals and
their organization in a goal stack are explained.

(2) Performance
This section covers the central subsymbolic concepts of ACT-R and discusses
the processes of activation spreading, partial matching, and conflict
resolution.  Empirical results and example models are presented to
illustrate and justify these concepts.

(3) Learning
ACT-R is a learning architecture which proposes mechanisms to learn new
symbolic knowledge (chunks, productions) as well as their subsymbolic
parameters (chunk and production strength, activation links, cost and
success estimates of productions, etc.).  In this section the various
learning mechanisms are introduced and illustrated by example
applications.

(4) Applications
Finally some ACT-R models and their relationship to empirical data are
presented in detail.  These models, varying from sequence learning to complex
problem solving, exemplify the scope of the ACT-R architecture and
emphazise its empirical success.  At the end of the tutorial we briefly
introduce ACT-R/PM, a newly developed perceptual-motor interface to
ACT-R.

Although we expect attendees to have some familiarity with the concept
of cognitive modelling, the tutorial will aim to be as accessible as
possible and the format will provide ample opportunity for discussions
and questions.