Metacognition in Computation: A selected research history and summary

1
Metacognition in Computation A selected
research history and summary

• Michael T. Cox
• BBNT Cambridge

2
Why Metacognition?

• What then can be the purport of the injunction,
  know thyself? I suppose it is that the mind
  should reflect upon itself.
• -- Augustine, De Trinitate, 16th century

3
Why Metacognition?

• Separates us from the rest of the species
• Separates smarter people from less smart
• Provides a heuristic basis for decisions
• E.g., I am good at home repair, so I can risk
  embarrassment by volunteering to fix the broken
  pipe rather than calling a plumber.

4
Metacognition is Ubiquitous
5
Why NOT Metacognition?

• Complexity space and time
• Actual human limitations
• Easier to show when metacognition does not work
  rather than how it does
• AI hype

6
AI Hype

• Once self-description is a reality, the next
  logical step is self-modification. Small,
  self-modifying, automatic programming systems
  have existed for a decade some large programs
  that modify themselves in very small ways also
  exist and the first large fully self-describing
  and self-modifying programs are being built just
  now. The capability of machines have finally
  exceeded human cognitive capabilities in this
  dimension it is now worth supplying and using
  meta-knowledge in large expert systems.
• -- Lenat, Davis, Doyle, Genesereth, Goldstein,
  and Schrobe 1983 (p. 238)

7
AI Hype

• Once self-description is a reality, the next
  logical step is self-modification. Small,
  self-modifying, automatic programming systems
  have existed for a decade some large programs
  that modify themselves in very small ways also
  exist and the first large fully self-describing
  and self-modifying programs are being built just
  now. The capability of machines have finally
  exceeded human cognitive capabilities in this
  dimension it is now worth supplying and using
  meta-knowledge in large expert systems.
• -- Lenat, Davis, Doyle, Genesereth, Goldstein,
  and Schrobe 1983 (p. 238)

8
What is Metacognition?

• Meta-X is defined as X about X
• Metacognition is cognition about cognition
• Metareasoning is reasoning about reasoning
• Metaknowledge is knowledge about knowledge
• Metamemory, metarepresentation,
  metacomprehension, metalogic, metaplans,meta...

9
But what about

• Meta-levels
• Reflection
• Introspection
• Self-awareness
• Self-explanation
• Consciousness?

10
Outline of Presentation

• Introduction, Motivation and Outline
• Early Roots
• Logic and Belief Introspection
• Knowledge-Based Systems, Metareasoning, and
  Control
• Case-Based Reasoning and Introspective Learning
• Conclusion

11
Setting the Stage

• Earliest AI Research Minsky McCarthy
• Minsky, M. L. 1965. Matter, Mind, and Models. In
  Proceedings of the International Federation of
  Information Processing Congress 1965 (Vol. 1)
  45-49.
• McCarthy, J. 1959. Programs with Common Sense. In
  Symposium Proceedings on Mechanisation of Thought
  Processes (Vol. 1), 77-84. London Her Majestys
  Stationary Office.
• Models of Models
• Declarative Knowledge for the Self

12
Minskys Theory

• Minsky, M. L. 1965. Matter, Mind, and Models. In
  Proceedings of the International Federation of
  Information Processing Congress 1965 (Vol. 1)
  45-49.
• To answer questions about the world and the self
  in the world, an agent must have a model it can
  query
• W, M, W, M, W, M

13
(No Transcript)
14
McCarthys Theory

• McCarthy, J. 1959. Programs with Common Sense. In
  Symposium Proceedings on Mechanisation of Thought
  Processes (Vol. 1), 77-84. London Her Majestys
  Stationary Office.
• Knowledge as logic
• Logic as thinking
• What does it mean for a robot to be conscious?

15
Logic Belief Introspection

• Self-Reference and aboutness (Perlis)
• Liars Paradox from time of Socrates
• This sentence is false.
• FOL axiomization and possible worlds (Moore)
• Belief is different than facts (Hintakka)
• Model-Theoretic reasoning
• Metalogics and proving provability

16
Konoliges Deduction Model

• Alternative to Possible Worlds Semantics
• Deduction Structure is a mathematical abstraction
  of bounded belief systems
• Machines and introspective machines
• Intrinsic and extrinsic self-beliefs
• Separation of IM from M resolves some problems of
  self-reference

17
Logical Representations

• Is-Complex-wrt
• (John,
• )
• How to handle?

18
Knowledge-Based Systems, Metareasoning Control

• Earliest Research Metaknowledge in expert
  systems
• Barr, A. 1977. Meta-Knowledge and Memory,
  Technical Report, HPP-77-37. Stanford University,
  Department of Computer Science, Stanford, CA.
• Davis, R. 1976. Applications of Meta-Level
  Knowledge to the Construction, Maintenance, and
  Use of Large Knowledge Bases. Stanford HPP Memo
  76-7. Stanford University.
• Metarules the red herring of AI

19
Davis Theory

• Knowledge engineering in MYCIN
• Metaknowledge
• Schemas
• Function templates
• Metarules
• Rule models
• Rule models help interpret what expert asserts

20
Example Rule Model

• INVESTMENT-AREA-IS
• Examples ((rule116 0.3) (rule050 0.7)
  (rule037 0.8) (rule095 0.9) (rule152 1.0))
• Description
• Premise ((returnrate same notsame 3.8)
  (timescale same notsame 3.8) (trend same)
  ((returnrate same)(timescale same) 3.8)
• Action ((investment-area conclude 4.7) (risk
  conclude 4.8))
• More-general (investment-area)
• More-specific (investment-area-is-utilities)

21
Model-based Understanding Learning by Experience
Expert
(dialog)
Knowledge Base
(knowledge acquisition)
Rule Acquisition
(concept formation)
(model-directed understanding)
Rule Models
22
Problems with Expert Systems

• Confuses abstraction with metacognition
• Confuses control with metacognition
• Self-understanding software tangent?
• Explanation is not a rule chain or proof tree
• Knows what it does not know?

23
Metareasoning

• Earliest Research Bounded rationality
• Simon, H. A. 1955. A Behavioral Model of Rational
  Choice. Quarterly Journal of Economics 69
  99-118.
• Earliest Research Goods Type II rationality
• Good, I. J. 1971. Twenty-Seven Principles of
  Rationality. In V. P. Godambe and D. A. Sprott
  eds. Foundations of Statistical Inference.
  Toronto Hold, Rinehart, Winston.

24
Wesfalds Theory

• Treating computation selection as action
  selection by maximizing expected utility
• Cost of time (world changes by itself)
• Benefit of better action choices
• Execution cost
• Resource cost

25
System Unification

• Unifies decision-making systems
• Decision-theoretic systems
• Production systems
• Goal-based systems
• Reactive systems
• EBL systems

• Unifies meta-cognitive systems
• MRS (Genesereth)
• TEIRESIAS (Davis)
• Soar (Newell)

26
Decision Stages and Shortcuts
A
Condition (s)
B
Condition (result(a,s))
E
C
F
Utility (result(a,s), v)
DT
D
Best (a,s)
27
Case-based Reasoning and Introspective Learning

• Earliest Research Schanks emphasis on memory
  and representation
• Schank, R. C., Goldman, N., Rieger, C., and
  Riesbeck, C. K. 1972. Primitive Concepts
  Underlying Verbs of Thought (Stanford Artificial
  Intelligence Project Memo No. 162. Stanford, CA
  Stanford University, Computer Science Department.
  (NTIS No. AD744634)
• Using Conceptual Dependency primitives to
  represent remember, forget, think, expect

28
Case-Based Explanation

• Provides a framework for interpreting Failures
• In world actions
• In reasoning actions (e.g., memory retrieval)
• In social actions
• Example Dog barking story
• S1 Police Dog enter airport baggage area
• S2 Dog sniffs luggage.
• S3 Dog Barks at luggage.
• S4 Police arrests suspect.

29
Computational Introspection

• To reason about the self
• When reasoning about the world fails use
  meta-reasoning to explain the failure
• Map from symptom of the failureto the cause of
  the failure
• Learn

30
Symptoms of Failure
Expectation does not exist
Expectationexists
Actual event exists
Impasse Surprise
Contradiction Unexpected Success
FalseExpectation
Actual eventdoes not exist
31
Causes of Failure
32
Stranded Motorist Example

• Planning a vacation
• Destination
• Reservation
• Supplies
• Gas
• Plan Execution
• Goes to store
• Buys supplies
• Drives to mountains
• Runs out of gas
• Failure Recovery
• Get gas can
• Walk to gas station or hitch-hike
• Fill can with gas
• Return to Fill tank
• Continue
• Failure Repair
• Regress goals to features in initial state
• Use features as index to store as new case

Causal Possibilities Tree

• Sub-goals
• Be at store
• Make purchases

33
Causal Possibilities Tree
X
34
Forgetting to Fill-Up with Gas
35
ConclusionProblems with Metacognition

• Control and rules
• Abstract rules are not really different from
  concrete rules
• Metaknowledge
• Knowledge about facts is not really different
  from ordinary facts
• Many synonymous and overloaded terms
• Overstating Benefits

36
Conclusion

• The Many-headed Monster of obscure parentage
  Brown (1987)
• Lessons to be learned
• Failures to be avoided
• Current research has the potential to be
  qualitatively different because of the technical
  maturities and funding commitments

37
A Grain of Salt

• To know oneself is only half as good as knowing
  two selves.
• --Homer
• (Simpson)

38
Knowing without Remembering