Astrolabe: A Robust and Scalable Technology - Cornell University
Envisioning a Robust, Scalable Metacognitive Architecture...
21
Envisioning a Robust, Scalable Metacognitive Architecture Built on Dimensionality Reduction Scruffy Metacognition Jason B. Alonso 1 Kenneth C. Arnold 2,3 Catherine Havasi 3,2 1 Personal Robots Group MIT Media Laboratory 2 MIT Mind Machine Project 3 Software Agents Group MIT Media Laboratory AAAI-10 Workshop on Metacognition for Robust Social Systems July 2010 Alonso, Arnold, Havasi (MIT) Scruffy Metacognition Metacognition 2010 1 / 21
Transcript of Envisioning a Robust, Scalable Metacognitive Architecture...
Envisioning a Robust, Scalable MetacognitiveArchitecture Built on Dimensionality Reduction
Scruffy Metacognition
Jason B. Alonso1 Kenneth C. Arnold2,3 Catherine Havasi3,2
1Personal Robots GroupMIT Media Laboratory
2MIT Mind Machine Project
3Software Agents GroupMIT Media Laboratory
AAAI-10 Workshop on Metacognition for Robust Social SystemsJuly 2010
Alonso, Arnold, Havasi (MIT) Scruffy Metacognition Metacognition 2010 1 / 21
Main Points
The Challenge
Metacognitive architectures too expensive to explore iterativelyProgrammers never count to 3.
Alonso, Arnold, Havasi (MIT) Scruffy Metacognition Metacognition 2010 2 / 21
Main Points
What to do?
How do we build a scalable framework for metacognitive architectures?
Hand-crafted metareasoners are out.Use many instances of a few types of simple but powerfulreasoning units.
Difference between reasoner and metareasoner in the inputsCommunicate with simple symbols, generally opaque semantics
What kinds of symbols? More later.
Alonso, Arnold, Havasi (MIT) Scruffy Metacognition Metacognition 2010 3 / 21
Main Points
What “reasoning units”?
What function should each component perform?Connectionist answer (switches, or “neurons”) theoreticallysatisfying to some, practically less than enlighteningOur answer: pattern discovery and matching
One basic process of an intelligent system is to identify usefulpatterns in its input and its outputOne symbol ⇐⇒ one pattern
Summarizing many inputs and outputs with fewer symbols... in essence, dimensionality reduction
Alonso, Arnold, Havasi (MIT) Scruffy Metacognition Metacognition 2010 4 / 21
Main Points
We claim...
Planning can be a pattern completion problem that leveragesdimensionality reductionMetacognitive functions, particularly metaplanning, can be built onthese principles
Alonso, Arnold, Havasi (MIT) Scruffy Metacognition Metacognition 2010 5 / 21
Vantage Point in AI
We are “Scruffies”
Imprecise and loosely statistical handling of symbolicrepresentations
Alonso, Arnold, Havasi (MIT) Scruffy Metacognition Metacognition 2010 6 / 21
Vantage Point in AI Corpora
We are “Scruffies”
Imprecise and loosely statistical handling of symbolicrepresentations
Open Mind Common Sense
Alonso, Arnold, Havasi (MIT) Scruffy Metacognition Metacognition 2010 7 / 21
Vantage Point in AI Corpora
We are “Scruffies”
Imprecise and loosely statistical handling of symbolicrepresentations
ConceptNet
Alonso, Arnold, Havasi (MIT) Scruffy Metacognition Metacognition 2010 8 / 21
Vantage Point in AI Analysis
We are “Scruffies”
Imprecise and loosely statistical handling of symbolicrepresentations
AnalogySpace, prepared
Alonso, Arnold, Havasi (MIT) Scruffy Metacognition Metacognition 2010 9 / 21
Vantage Point in AI Analysis
We are “Scruffies”
Imprecise and loosely statistical handling of symbolicrepresentations (dynamic representations)
AnalogySpace, computed
Alonso, Arnold, Havasi (MIT) Scruffy Metacognition Metacognition 2010 10 / 21
Scruffy Metacognition An example
Sheep games
“Nexi, come take the sheep.”
Alonso, Arnold, Havasi (MIT) Scruffy Metacognition Metacognition 2010 11 / 21
Scruffy Metacognition An example
Sheep games, continued
Now imagine two games...Both involve picking up toysand putting them somewhereHow could Nexi know whichgame we’re playing?
Alonso, Arnold, Havasi (MIT) Scruffy Metacognition Metacognition 2010 12 / 21
Scruffy Metacognition An example
An architecture
Metaplanner
EN CON
V
EN CON
V
SensorySystem
MotorSystem
Action
Observation
EN Planner Enable
Planner ConvergenceCO
NV
Multiple symbols
Alonso, Arnold, Havasi (MIT) Scruffy Metacognition Metacognition 2010 13 / 21
Scruffy Metacognition Components
The Reducer
Mixed inputsDimensionality-reducedoutputs
Candid, Covariance-free Incremental Principal ComponentAnalysis (CCIPCA) (Weng et al. 2003)Essentially AnalogySpace
Dynamically-generated representationsOpen domain
Alonso, Arnold, Havasi (MIT) Scruffy Metacognition Metacognition 2010 14 / 21
Scruffy Metacognition Components
The Planner, part 1 of 2
Past Present Future
Build a model of salient patterns inobservable events and behaviorsGenerate plans that achieve goalsgiven this modelIncremental. Learn/refine modelsfrom experience in real timeScruffy. Statistical handling ofsymbolic representations of the realworld to draw robust conclusionsIn practice, two approaches:
Replay of natural responses toenvironment and teammatesGoal-seeking
Alonso, Arnold, Havasi (MIT) Scruffy Metacognition Metacognition 2010 15 / 21
Scruffy Metacognition Components
The Planner, part 2 of 2
Bus of observations
Single action
Reduced
timelineAction
Observation
EN Planner Enable
Planner ConvergenceCO
NV
Multiple symbols
EN CON
V
Alonso, Arnold, Havasi (MIT) Scruffy Metacognition Metacognition 2010 16 / 21
Scruffy Metacognition Components
Intuitive Introduction to MIDAWT
A
B
Ctime
weight
time
~f [t ] = (∑
i ai ∗ ~gi)[t ]
Given a multivariate waveform(timeline):
Detect instances ofpreviously-seen patternsRefine models for thosepatterns (or record newpattern)Describe timeline as acombination of understoodpatternsComplete timeline byinterpolating gaps in timeline
Alonso, Arnold, Havasi (MIT) Scruffy Metacognition Metacognition 2010 17 / 21
Scruffy Metacognition An example, revisited
An architecture
Metaplanner
EN CON
V
EN CON
V
Sensory
System
Motor
System
Action
Observation
EN Planner Enable
Planner ConvergenceCO
NV
Multiple symbols
Alonso, Arnold, Havasi (MIT) Scruffy Metacognition Metacognition 2010 18 / 21
Summary
Summary
The difference between cognition and metacognition is in thewiring, permitting scalable architectures.Systems that build their own representations dynamically aremore robust.
Forthcoming experimentsMars EscapeRestaurant GameExplore/exploit
Alonso, Arnold, Havasi (MIT) Scruffy Metacognition Metacognition 2010 19 / 21
Extra slides
Here be dragons
Alonso, Arnold, Havasi (MIT) Scruffy Metacognition Metacognition 2010 20 / 21
Extra slides
A Unique MIDAWT/SP Insight
When the astronaut player is engaged in a search activity with abunch of boxes, the robot is not about to hit the elevator callbutton.Correlation not causal, but reflective of teaming behaviorAnti-correlation not found in CBP or plan networks
Alonso, Arnold, Havasi (MIT) Scruffy Metacognition Metacognition 2010 21 / 21
