ACT-R Workshop John R. Anderson Daniel Bothell Christian Lebiere Niels A. Taatgen
description
Transcript of ACT-R Workshop John R. Anderson Daniel Bothell Christian Lebiere Niels A. Taatgen
ACT-R Workshop
John R. AndersonDaniel Bothell
Christian LebiereNiels A. Taatgen
Schedule of Events: 9:00-10:30: ACT-R from CMU’s Perspective 11:00-12:30: Architecture 1:30-3:30: Extensions 4:00-5:30: Future of ACT-R from a non-CMU Perspective
And lots of Interaction!
ACT-R Workshop Schedule
Opening: ACT-R from CMU’s Perspective9:00 - 9:45 Overview of ACT-R -- John R. Anderson9:45 – 10:30 Details of ACT-R 6.0 -- Dan Bothell
Break: 10:30 – 11:00Presentations 1: Architecture
11:00 – 11:30 Functional constraints on architectural mechanisms -- Christian Lebiere11:30 – 12:00 Retrieval by Accumulating Evidence in ACT-R -- Leendert van Maanen12:00 – 12:30 A mechanism for decisions in the absence of prior reward -- Vladislav D. Veksler
Lunch: 12:30 – 1:30Presentations 2: Extensions
1:30 – 2:00 ACT-R forays into the semantic web -- Lael J. Schooler2:00 – 2:30 Making Models Tired: A Module for Fatigue -- Glenn F. Gunzelmann2:30 – 3:00 Acting outside the box: Truly embodied ACT-R -- Anthony Harrison3:00 - 3:30 Interfacing ACT-R with different types of environments and with different techniques: Issues and Suggestions.-- Michael J. Schoelles
Break: 3:30 – 4:00Panel: 4:00 – 5:30: Future of ACT-R from a non-CMU Perspective Danilo Fum, Kevin A. Gluck, Wayne D. Gray, Niels A. Taatgen, J. Gregory Trafton, Richard M. Young
Overview of ACT-R
John R. AndersonCarnegie Mellon University
Outline:
9:10: Big picture of what ACT-R is about
9:20: Evolution of the Procedural Module
9:30: Evolution of the Declarative Module
9:35: How ACT-R spreads
ACT-R is Not Monolithic
1. It is a community brought together by common theoretical assumptions and a commitment to the “No Magic” Principle -- cognitive theory has to run and it has to predict data. While ACT-R may be sustained from CMU it no longer resides at CMU. The community motto is “Let a thousand flowers grow”
2. It is a set of software for purposes of simulation. This software consists of a core LISP implementation, but there are many theoretically-motivated extensions and alternative practicality-motivated alternative implementations. In some cases the software provides the best definitions of what the theoretical claims are.
3. It is a theory that attempts to formalize and operationalize certain aspects of our understanding of the human mind. This includes assumptions that are more core and those that are more peripheral. It changes as our knowledge grows and has different interpretations in different hands.
ACT-R: The Oldest Core Principles
1. The Procedural-Declarative Distinctiona. The declarative component originated in Anderson & Bower (1973)
HAM network representation of memory.b. The procedural component originated in Newell’s (1973) production
system theory of cognitive control.c. Both the procedural and declarative components have evolved far
from these origins.2. The Symbolic-Subsymbolic Distinction
a. In addition to the symbolic level that represented knowledge there is a subsymbolic level that controls access to that knowledge.
b. The subsymbolic level was initially designed to reflect the 1970s & 1980s ideas about neural processing.
c. Guided by rational analysis the subsymbolic level was updated in 1993 to reflected the likelihood that the information was useful. This was the birth of ACT-R.
Evolution from ACT-R 2.0 (1993) to ACT-R 6.0 (2007)
1. There were 3 driving forces:a. The emergence of a user community around the publicly available
ACT-R 2.0.b. The realization that the “No Magic” principle required that we be
able to model the processing all the way from input to output. c. The insistence on not making assumptions that could not be
cashed out into neurally plausible computations.2. This converged in the modular architecture of ACT-R 6.0:
a. The allowed community members to try variations on existing ideas and extensions but keep what they wanted.
b. We borrowed the modular organization of EPIC for the perceptual-motor modules.
c. There was growing evidence that, while the brain was a complex parallel machine, different regions had their specializations.
Modules are high capacity, parallel, and asynchronous
Modules in ACT-R 6.0
Production system that contains rules that recognize patterns and react
Buffers provide narrow paths of communication -- only hold a chunk in ACT-R terms.
Manual Vocal
Visual Aural
Imaginal Declarative
Goal
Procedural
ACT-R Module-Region Mappings
Outline:
9:10: Big picture of what ACT-R is about
9:20: Evolution of the Procedural Module
9:30: Evolution of the Declarative Module
9:35: How ACT-R spreads
The Procedural Component in ACT-R has Evolved from Computer Science Notation to Description of
the Brain’s Action Selection
600517- 23523 4
The First Real ACT-R Production RuleIf the goal is to process a column and the top digit is not smaller than the bottom digit,Then write the difference between the digits as the answer
Responds to a Particular Pattern that Appears in the Buffers of a Set of Modules
Which consists of requests to other Modules
Imaginal> Top: 7 Relation: >= Bottom: 3
Goal> Task: Process-Column
Declarative> Type: subtraction Minuend: 7 Subtrahend: 3
Goal> Task: Subtracting
RequestDifference
Selects anAction
The Second Real ACT-R Production RuleIf the goal is to process a column and the top digit is not smaller than the bottom digit,Then write the difference between the digits as the answer
Goal> Task: Subtracting
Declarative>Type: subtractionDifference: 4
Manual> Action: write Digit: 4
Goal> Task: Next-column
Harvest Difference
Responds to a Particular Pattern that Appears in the the Buffers of a Set of Modules
Which consists of requests to other Modules
Selects anAction
Attributes of Production Rules
Production rules are stimulus-response bonds that have “gone over to the cognitive side” because among the stimuli they respond to are past memories, mental images, and control states.
Respond to conjunctions of elements in the various buffers. These buffers can represent relational structures -- e.g. A
above B. Note how innocuous the use of variables is -- it basically
copying information from one brain region to another.
Stewart, T.C. and Eliasmith, C. (2008). Building production systems with realistic spiking neurons. 30th Annual Meeting of the Cognitive Science Society.
Stocco, A., Lebiere, C., & Anderson, J. R. (in revision). Conditional routing of information to the cortex: A model of the role of basal ganglia in high-level cognition. Psychological Review
New Problem Situations
Following instructions (e.g. Multicolumn Subtraction)
Declarative Representations
Requires Deliberation
Analogy to Prior Experiences (e.g. Past Tense Model)
Interpreted
New Production Rules
EventuallyProduces
Production Compilation
Traces Feed Into
Learning of New Production Rules
DeductionFrom 1st Principles
Imaginal> Relation:
Goal> Task: Process-Column State: Imaginal
RetrieveOperator
Retrieval> Type: operator Pre:
Imaginal> Top: Bottom:
PerformSubtraction
Retrieval> relation: subtract arg1: top arg2: bottom post:
Retrieval> type: subtraction minuend: subtrahend:
Goal> Task:
StateFeature
StateFeature
ActionIdentity
ObjectReferent
ObjectReferent
ActionType
Operator
Pre
Post
arg2
Arg1
Action
Type
Op11-1
Top >= Bottom
Subtract
Retrieve
Bottom
Top
Subtracting
Imaginal> Top: Relation: >= Bottom:
Goal> Task: Process-Column
RequestDifference
Retrieval> Type: subtraction Minuend: Subtrahend:
Goal> Task: Subtracting
Origin of One of the Subtraction Rules
Production compilationcompresses general-purpose processing of knowledge into special case rules -- replacing deliberation by action.
Reinforcement of Competing Productions
Retrieve-Instruction (Reinforcement 10)If the goal is to process a columnThen retrieve an operator for that kind of column
Request-Difference-Subtract (Reinforcement 14)If the goal is to process a column and the top digit is not smaller than the bottom digit,Then subtract the bottom from the top
Request-Difference-Wrong (Reinforcement 14 or 0)If the goal is to process a columnThen subtract the smaller from the larger
Request-Difference-Borrow (Reinforcement 14)If the goal is to process a column and the top digit smaller than the bottom digit,Then add 10 to the top digit and set as a subgoal to borrow from the column to the left.
Utility Learning for Competing Productions
00 .10 .20 .30 .40 .50 .60 .70 .80 .9102 55 07 51 0 0E x p e rie n c e sIn s tru c tio nS u b tra c tB o rro wW ro n g
€
U i(n) =U i(n −1) +α [Ri(n) −U i(n −1)]
Considerable simplification of ACT-R utility learning based of reinforcement-like learning results from the basal ganglia
0
2
4
6
8
10
12
14
0 25 50 75 100Experiences
InstructionSubtractBorrowWrong
€
Pi =eUi / s
eUj / s
j
∑Standard ACT-R soft-max rule for choosing among productions according to their noisy utilities0
0.10.20.30.40.50.60.70.80.9
1
0 25 50 75 100
Experiences
InstructionSubtractBorrowWrong
Every time a rule created it is rewarded with the utility of its parent
Outline:
9:10: Big picture of what ACT-R is about
9:20: Evolution of the Procedural Module
9:30: Evolution of the Declarative Module
9:35: How ACT-R spreads
What has Happened to the Declarative Component in ACT-R?
It has bifurcated into two completely separate things:
1. An increasingly watered-down set of principles for the representation of knowledge, which comes to be the contents of module buffers. This is clearly a place where important new thinking is required.
2. An increasingly empirically well-founded set of principles (with a foundation in rational analysis) for how the brain performs controlled retrieval of information from declarative memory.
Buffers and Declarative Memory Buffers associated
with modules provide narrow paths of communication.
The contents of the buffers are called chunks.
Records of these chunks are placed in declarative memory.
These can be later retrieved and placed in the declarative buffer.
Manual Vocal
Visual Aural
Imaginal Declarative
Goal
Procedural
Chunk Activation Reflects Probability of Use
€
Log(Posterior(i |C)) = Log(Pr ior(i)) + Log(Likelihood( j | i)j∈C
∑ )Environmental Equation:
€
Ai = Bi + W jS ji
j∈C
∑Activation Equation:
Base-level Activation of memory i Association Strength from j to i
Posterior odds that memory i will be needed in context C
Likelihood ratio of element j in context given i is needed
Prior odds that i is needed: recency and frequency
Momentary Activation of memory i Weighting of Source j
Fan Experiment: Pirolli & Anderson (1985)
€
S ji = S − ln(Fan)
Growth of Activation
Ai = Bi + ΣWj S ji
0
0.5
1
1.5
2
2.5
1 2 3 4 5 6 7 8 9 10 Days of Practice
Activation Level1-1 Fan3-1 Fan3-3 Fan
Act
ivati
on L
evel
Recognition Latencies
0
200
400
600
800
1000
1200
1400
1 2 3 4 5 6 7 8 9 10
Days of Practice
Recognition Time (ms.)
3-3 Fan3-1 Fan1-1 FanSeries4Series5Series6
Reco
gnit
ion T
ime
(ms.
)
r = .986 is a parameter-free measure of the match between theory and data.
€
Time = I + Fe−A i Re trieval Time Equationintercept latency scale
Outline:
9:10: Big picture of what ACT-R is about
9:20: Evolution of the Procedural Module
9:30: Evolution of the Declarative Module
9:35: How ACT-R spreads
Temporal Module: An Example of How One Can Extend ACT-R
PacemakerPacemaker GateGate
StartSignalStart
Signal
AccumulatorAccumulator
MemoryMemory
ComparisonComparison
MatchingMatching
SelectionSelection
ExecutionExecutionPro
du
ctio
ns
Pro
du
ctio
ns
Declarative ModuleDeclarative Module
Visual ModuleVisual Module Manual ModuleManual Module
External WorldExternal World
Retrieval BufferRetrieval Buffer
Manual BufferManual BufferVisual BufferVisual Buffer
Goal BufferGoal Buffer
Problem BufferProblem BufferPacemakerPacemaker GateGate
StartSignalStart
Signal
AccumulatorAccumulator
Other Module Extensions for ACT-R
Salvucci’s Emma Module for Eye Movements.
My new Metacognitive Module.
Spatial Modules (Gunzelmann, Harrison & Trafton).
Fatigue Module (Gunzelmann) ????
Reasoning Module LarKC (Schooler)????
Module ModificationsSNIF-ACT (Fu & Pirolli): Procedural and Declarative.
Threaded Cognition (Salvucci & Taatgen): Goal
Spacing Effect (Pavlik): Declarative.
Blending (Lebiere): Declarative.
Race/A (van Maanen & Van Rijn): Declarative
Visual Saliency (Byrne): Visual.
Gray, Veksler, & and others of the RPI Co: Procedural.
Bothell & Leabra: Visual.
You Don’t Need to Change ACT-R to Have an Interesting Model
Fum & Stocco: Sugar Factory
Lebiere, Wallach, & Taatgen: Sugar Factory
Altmann & Trafton: Tower of Hanoi
Lewis & Vasishith: Parsing
Taatgen: Acquistion Past Tense Model
Anderson (2007) & Everybody (recently): Everything in fMRI
And indeed most of the published ACT-R models.
Getting ACT-R out of the Narrow Confines of Laboratory Experiments
Best & Lebiere: MOUT
St. Amant & Ritter: Segman
Bothell, Douglass, Lee: Unreal Tournament
Harrison & Trafton: Robotics
Destefano: Space Fortress
Schoelles: Lots of Interfaces
ACT-R is Not Monolithic
1. It is a community brought together by common theoretical assumptions and a commitment to the “No Magic” Principle -- cognitive theory has to run and it has to predict data. While ACT-R may be sustained from CMU it no longer resides at CMU. The community motto is “Let a thousand flowers grow”
2. It is a set of software for purposes of simulation. This software consists of a core LISP implementation, but there are many theoretically-motivated extensions and alternative practicality-motivated alternative implementations. In some cases the software provides the best definitions of what the theoretical claims are.
3. It is a theory that attempts to formalize and operationalize certain aspects of our understanding of the human mind. This includes assumptions that are more core and those that are more peripheral. It changes as our knowledge grows and has different interpretations in different hands.
Be Fruitful and Multiply!
(p. 12 Architecture of Cognition, 1983) 2007