MIDCA: A Metacognitive, Integrated Dual-Cycle Architecture for
Toward an Integrated Metacognitive Architecture
description
Transcript of Toward an Integrated Metacognitive Architecture
Cox – 8 July 2011
MICHAEL T. COXUMIACS, UNIVERSITY OF MARYLAND, COLLEGE PARK
Toward an Integrated Metacognitive Architecture
http://xkcd.com/
Cox – 8 July 2011
2
Why a Metacognitive Architecture?
Why Cognitive Architectures? To better understand the mechanisms of reasoning across
tasks To account for human data To study high-level cognition by specifying the underlying
infrastructureMetacognition because it is especially human and
gets at the nature of what it means to be intelligentIntegrated because many different aspects exist
And much of it is confused And none have put it all together And this is the only way to get at human-level AI
Cox – 8 July 2011
3
INTRODUCTIONOUTLINE
COGNI TI VE AND METACOGNI TI VE ARCHI TECTURES
REPRESENTATI ONSTHE SELF-REGULATED LEARNI NG TASK
CONCLUSI ON
Outline
Cox – 8 July 2011
4
INTRODUC TIONOUTLI NE
COGNITIVE AND METACOGNITIVE ARCHITECTURES
REPRESENTATI ONSTHE SELF-REGULATED LEARNI NG TASK
CONCLUSI ON
Cognitive and Metacognitive Architectures
5
Action and Perception Cycle
Doing Reasoning
from Russell & Norvig, 2002
Cox – 8 July 2011
6
ObjectLevel Meta-Level
GroundLevel
Doing Reasoning Metareasoning
ActionSelection
Meta-level Control
PerceptionIntrospectiveMonitoring
Simple Model of Metareasoning
ObjectLevel Meta-Level
GroundLevel
Doing Reasoning Metareasoning
ActionSelection
Meta-level Control
PerceptionIntrospectiveMonitoring
from Cox & Raja (2011)
Cox – 8 July 2011
7
The Meta-Cognitive Loop (MCL)
indications failures responses
expectations correctionsMCL
host system
abst
ract
conc
rete
Meta-level ControlIntrospective Monitoring
Cox – 8 July 2011
from Anderson et al., (2008)
8
Meta-AQUA Metacognitive Architecture
TaleSpin
MultistrategyStory
Understanding
MultistrategyLearning
CBR PlannerG
Story Representation
FK
BK
Learning Algorithm Toolbox
Trace
Learning Plans
Execute Learning
caselibrary
XPLibrary
scriptlibrary
is-ahierarchy
Performance SubsystemProblem Generation
Learning Subsystem
Story Input
MemoryLearning
Algorithms
Learning Goals
XPs
∆BK
Introspective Monitoring
Meta-levelControl
Cox – 8 July 2011
from Cox & Ram (1999)
Cox – 8 July 2011
9
INTRO: The INitial inTROspective Agent
Ground Level
Object Level
Object Level
Object and Meta-Level
from Cox (2007)
10
Cognitive Model
Domain
Intend
Act (& Speak)
Plan
Evaluate
Perceive (& Listen)
Interpret
Goals
from Norman (1986)
Cox – 8 July 2011
MemoryWorld Model
Plans
Semantic Memory
Episodic Memory
Visual Memory
goal change goal inputresolve
anomalygoal
subgoal
ProblemSolving
Explanation
11
Metacognitive Model
Cox – 8 July 2011
7Mental Domain
goal change goal input
Intend
Control
Plan
Evaluate
Monitor
Interpret
Meta Goals
Meta-LevelControl
Introspective Monitoring
subgoal
MemoryReasoning Trace
Strategies
Episodic Memory
Metaknowledge
Self Model
resolve anomaly
goal
Cox – 8 July 2011 12
An Integrated Metacognitive Architecture
Domain
MemoryWorld Model
Plans
Semantic Memory
Episodic Memory
Visual Memory
ProblemSolving
Explanation
goal change goal inputresolve
anomalygoal
Intend
Act (& Speak)
Plan
Evaluate
Perceive (& Listen)
Interpret
Goalssubgoal
4Mental Domain
Intend
Control
Plan
Evaluate
Monitor
Interpret
Meta Goals
Meta-LevelControl
Introspective Monitoring
subgoal
MemoryReasoning Trace
Strategies
Episodic Memory
Metaknowledge
Self Model
resolve anomaly
goal
Goal Managementgoal change goal input
Cognition
Metacognition
Cox – 8 July 2011
13
INTRODUC TIONOUTLI NE
COGNIT I VE AND METAC OGNIT I VE ARC HI TECTURESREPRESENTATIONS
THE SELF-REGULATED LEARNI NG TASKCONCLUSI ON
Representations
Cox – 8 July 2011
14
Representations For Mental Traces
15
Truth Values on Graph Nodes
Cox – 8 July 2011
Description
A E G I M
Absent Memory
inFK outFK inFK outBK outBK
Absent Index inFK outFK inFK outBK inBK
Absent Question
inFK outFK outFK x x
Absent Feedback
outFK outFK x x x
X=don’t care
16
Partial Ontology for Mental Terms
Cox – 8 July 2011
17
Self-Models
Cox – 8 July 2011
How to represent episodic memory? Case-based reasoning Soar’s episodic memory
How to represent model of self? Physical attributes Mental attributes
Dispositions Attitudes Emotions Intellectual abilities
Social attributes
Cox – 8 July 2011
18
INTRODUC TIONOUTLI NE
COGNIT I VE AND METAC OGNIT I VE ARC HI TECTURESREPRESENTATI ONS
THE SELF-REGULATED LEARNING TASKCONCLUSI ON
The Self-Regulated Learning Task
Cox – 8 July 2011
19
Task: Self-Regulated Learning (SRL)
SRL focuses on deliberate learningSRL scope is wide and task is difficultSRL has extant data (e.g., Azevedo)The problem of studying for a test
Must master the domain Must understand one’s self
One’s own knowledge One’s own reasoning ability
Must understand the teacher’s priorities
Cox – 8 July 2011
20
How to Study for a Test
Reason about the domain (e.g., chemistry)Reason about one’s knowledge of the domainReason about skills in the domain (e.g., lab skills)Reason about reasoning (problem-solving) in the
domainReason about personal strengths and weaknesses
in domain (I struggled with Chem I, so need to work harder; I study best in quiet environments)
Reason about teacher and what is likely to be on test
Reason about resources (e.g., time left to study)
Cox – 8 July 2011
21
Task Decomposition I
ContextReading assignment,
take notesAttend lecture, take
notesPerform homeworkStudy for testTake test
Study for testReview notesReview readingsReview old testsPractice problems
Cox – 8 July 2011
22
Task Decomposition II
To review readingsMust have indicated
key parts when first read
Integrate notes from lecture
Identify parts needing elaboration
Do elaborationIterate until confident
or no time remaining
Lecture
NotesBasic backgroundKey textKey textPartially understoodPartially understood
Figure Caption
Figure
Homework
Readings
Teacher ModelSelf Model
Time left ¬
prepared?
yesno Halt
Cox – 8 July 2011
23
Desiderata
System that has self-identity Knows its own strengths and weaknesses Knows what it does not know Knows what it wants for the future Has a memory for what it has done in the past Has a sense of its current physical presence in space
and time (e.g., knows what is graspable) Is self-confident and acts deliberately Can empathize with others Can explain itself to others Generates its own goals (is an independent actor) *Wonders about what happens when it gets turned off
Cox – 8 July 2011
24
Self-Description
Cox – 8 July 2011
25
INTRODUC TIONOUTLI NE
COGNIT I VE AND METAC OGNIT I VE ARC HI TECTURESREPRESENTATI ONS
THE SELF-REGULATED LEARNI NG TASKCONCLUSION
Conclusion
Cox – 8 July 2011
26
Conclusion
A number of different architectures exist that bear on metacognition
None have integrated the many aspects of cognition and metacognition
To do so would capture something uniquely human and at the heart of what it means to be intelligent
This presentation represents a small start