CogSci C131
Production systems and
cognitive architectures
Tom Griffiths
Admin
Problem set 0 is due 5pm Friday!
Python transition class 4:30 today in
2305 Tolman
All sections meet as normal this week
Office hours are now on bCourses
Reader is available from Fast Imaging
(2022 University) for $35.60+tax
Token manipulation systems
System is defined fully by
a set of tokens
starting positions for those tokens
formal rules stating how token positions can
be changed into other token positions
Rules depend only on current positions,
and define only the next positions
Example 2: Formal logic
Pieces
P, Q, , , , , (, )
Starting positions
well-formed formulas
e.g., P Q
Formal rules
e.g., P Q
P
Q
Outline
Production systems
Break
Cognitive architectures
Logic
Inference Rules
Facts
PQ
Q
P
The World
Logic Theorist
(Newell & Simon, 1956)
Herb Simon and Allen Newell
The first AI system
Found proofs by using
heuristics to search
the large space
Heuristics inspired by
human problemsolving strategies
Symposium on Information Theory
Often considered the birth of cognitive science
(on 9/11/56, at MIT)
Three famous papers presented:
Allen Newell & Herbert Simon, The Logic Theory
Machine: A complex information processing system
Noam Chomsky, Three models of language
George Miller, The magical number seven
Logic
Inference Rules
Facts
PQ
Q
P
The World
Early AI systems
Operations
Workspace
Facts
Operations
Goals
Actions
Observations
The World
Production systems
Operations are encoded in IF THEN form
Production fires when IF condition is
satisfied, executes instructions following THEN
String World
The world is a string of characters
e.g. ABC
Facts are properties of the string
e.g. initial substrings
Actions are rewriting parts of the string
e.g. AB -> BA
A production system
P1:
P2:
P3:
P4:
P5:
P6:
$$ -> *
*$ -> *
*x -> x*
* -> null & halt
$xy -> y$x
Where:
null -> $
x and y can be any letter
* and $ are reserved symbols
null is the empty symbol
halt ends execution
Some challenges
How do we decide which production to apply,
when many are possible?
execute the first
execute at random
execute via some complicated criterion
Should we allow any IF THEN rule to be in
our set of rules?
string = computation;
disp(string);
thought
Minds and computers are both formal systems
ACT-R
(Anderson, 1993)
Application
Declarative
Memory
Production
Memory
Match
Storage
Retrieval
Working
Memory
Encoding
Execution
Performance
Outside World
John Anderson
Declarative memory
Facts are represented by chunks
Chunks are activated by context
organized in a semantic network, which we
will talk about later in the semester
Procedural memory
Skills and strategies used in solving
problems are represented as productions
Hierarchy of goals viewed as one of the
key characteristics of human behavior
Productions differ in time to execute
Learning
People can acquire new chunks through
perceptual experience, or via productions
People can form new productions by
generalizing from experience
Learning
People can acquire new chunks through
perceptual experience, or via productions
People can form new productions by
generalizing from experience
Sets of productions can be compiled into
more efficient productions
Compiling productions
If certain productions regularly fire other
productions, then those productions can be
collapsed
Can be executed by another production
The power-law of practice
time = practice-
1023 alternative
choice task
(Siebel, 1963)
log(time) = -log(practice)
Asimovs
books
(Ohlsson, 1992)
The promise of production systems
All that there is to intelligence is the simple
accrual and tuning of many small units of
knowledge that in total produce complex
cognition
(Anderson, 1996, p. 356)
Simons (1981) metaphor of the ant
Just a matter of more rules
Doug Lenat
Same philosophy has
been applied to AI
The Cyc project aims to
collect rules for common
sense reasoning
You can download a
version with 300,000
concepts, 2,000,000 facts
The promise of production systems
All that there is to intelligence is the simple
accrual and tuning of many small units of
knowledge that in total produce complex
cognition
(Anderson, 1996, p. 356)
Simons metaphor of the ant
Productions can modify productions
Early AI systems
Operations
Workspace
Facts
Operations
Goals
Actions
Observations
The World
The promise of production systems
All that there is to intelligence is the simple
accrual and tuning of many small units of
knowledge that in total produce complex
cognition
(Anderson, 1996, p. 356)
Simons (1981) metaphor of the ant
Productions can modify productions
A programming language for human thought
just a matter of figuring out the productions!
Chatbots and productions
Break
Up next:
Cognitive architectures
Computation
Alan Turing
(1912-1954)
Computer architecture
John von Neumann
What architecture underlies cognition?
How can we characterize human information
processing in terms of
memory
symbols
operations
interpretation
interaction with the world
Cashing out the idea of cognition as a formal
system, explicitly analogous to computers
ACT-R
(Anderson, 1993)
Application
Declarative
Memory
Production
Memory
Match
Storage
Retrieval
Working
Memory
Encoding
Execution
Performance
Outside World
John Anderson
SOAR
(Newell, Rosenbloom, Polk, and Laird, 1987)
Production Memory
Chunking
Execution
Working Memory
Perceptual
Systems
Motor
Systems
Senses
Muscles
Outside World
Decision
Allen Newell
Why study cognitive architecture?
Estimate parameters characterizing
human information processing
how long does each operation take?
Additive factors analysis
Assuming that each operation takes a
fixed amount of time, estimate times by
manipulating operations involved
H
P
Z
Probe: Q
M
Probe: L
N
S
H
P
A
R
V
C
Probe: B
Additive factors analysis
Assuming that each operation takes a
fixed amount of time, estimate times by
manipulating operations involved
(Sternberg, 1967)
So estimate time for productions, to
build up to really complex tasks
Why study cognitive architecture?
Estimate parameters characterizing
human information processing
how long does each operation take?
Figure out whats going wrong
Why study cognitive architecture?
Estimate parameters characterizing
human information processing
how long does each operation take?
Figure out whats going wrong
Unified theories of cognition
rather than lots of different models, lots of
programs in a common language
shared components across different
models, connections between problems
What kind of explanations do we get?
What sort of unified theory do we develop?
Marrs three levels
constrains
Computation
What is the goal of the computation, why is it
appropriate, and what is the logic of the strategy
by which it can be carried out?
constrains
Representation and algorithm
What is the representation for the input and
output, and the algorithm for the transformation?
Implementation
How can the representation and algorithm be
realized physically?
What kind of explanations do we get?
What sort of unified theory do we develop?
Even if production systems are the right
way to characterize human cognition, all
we get is a mechanistic story
a program for every problem
describes how people do what they do
not why people behave in a particular way
Thursday
Formal systems and language
you get to read Chomsky!
think about different proposals for formal systems
that can capture the structure of human languages
