Flexible Reasoning with Functional Models

1
Flexible Reasoning with Functional Models
J. William Murdock Intelligent Decision Aids
Group Navy Center for Applied Research in
Artificial Intelligence Naval Research
Laboratory, Code 5515 Washington, DC
20375 bill_at_murdocks.org http//bill.murdocks.org
Presentation at Ohio State University April 15,
2003
2
General Motivation

• Complex dynamic environments demand quick and
  flexible behavior.
• Specialized software is quick but not flexible.
• Generative planning, reinforcement learning, etc.
  are very flexible but very slow.
• Functional models provide the power of
  specialized software and the flexibility of AI.

3
Specific Objectives (Outline)

• Retrospective adaptation
• A system encounters a new constraint during
  execution
• It uses a model of itself to redesign itself.
• Proactive adaptation
• A system is told to perform a new, unknown task.
• It must redesign itself before it can do an
  execution.
• Models for existing, similar tasks may be
  adapted.
• Or new models may be built from scratch.
• Self-explanation
• A system uses models of its reasoning process and
  products to explain itself and justify its
  results.
• Explanation of threats
• A system has a model of enemy behavior.
• Uses that model to explain a potential threat and
  help a user decide whether that that threat is
  genuine.

4
TMK (Task-Method-Knowledge)

• TMK models encode knowledge about a process
  (e.g., a computer program, a military activity).
• TMK encodes
• Tasks functional specification / requirements
  and results
• Methods behavioral specification / composition
  and control
• Knowledge Domain concepts and relations

5
Partial History of TMK
Key Influences
Functional Representation (Sembugamoorthy
Chandrasekaran 1986)
Generic Tasks (Chandrasekaran 1986)
OSU
SBF Models (Goel, Bhatta, Stroulia 1997)
ZD (Liver and Allemang 1995)
TMK Projects
Autognostic Failure-driven learning (Stroulia
Goel 1995)
...
Interactive Kritik Self-Explanation (Goel
Murdock 1996)
GT
SIRRINE Retrospective Adaptation (Murdock Goel
2001)
REM Proactive Adaptation (Murdock 2001)
ToRQUE Scientific Cognition (Griffith 1999)
AHEAD Explanation of Threats (Murdock, Aha,
Breslow 2003)
NRL
DiscoveryMachine
6
SIRRINERetrospective Adaptation

• Self-Improving Reflective Reasoner Integrating
  Noteworthy Experience
• A shell for adaptive software systems.
• Systems encoded in SIRRINE contain a TMK model of
  themselves
• Used to automate adaptation in response to new
  constraints for a known task.

7
SIRRINE Functional Architecture
Model
8
Meeting Scheduling Agent

• Example agent which schedules a regular weekly
  meeting.
• Given a length of time and a list of schedules,
  it produces a time slot that fits into those
  schedules.
• It has a set of time constraints which require
  that meetings be held from 9AM to 5PM on Monday
  through Friday.

9
Diagrams of TMK Models
10
Model Animation
11
Example Problem

• Research group needs to schedule a 90 minute
  meeting. There are no available 90 minute slots
  between 9AM and 5PM on Mondays through Fridays.
• The agent fails for this problem. Feedback is
  given stating that the meeting should be held on
  Tuesdays from 430PM to 600PM.
• Credit assignment process identifies the
  find-next-slot task as one which could have
  produced the desired result.
• Modification process alters that task.

12
Knowledge forCredit Assignment

• Feedback State what the overall results should
  have been.
• Trace State what the results actually were also
  used to localize failure
• Models indicate differences between what the
  results should have been and what they were
  drive the modification process

13
Trace
14
Credit Assignment Process

• Heuristics guide search through trace
• Causal proximity temporally closest to end first
• Functional abstraction most abstract first
• Model used at each step of the search
• When a potential contradiction is found, a
  particular type of credit is assigned
• task-does-not-produce-value, method-does-not-produ
  ce-value, primitive-generates-invalid-state,
  primitive-fails
• Result Localization of credit

15
Modification

• Library of modification strategies indexed by the
  type of credit assigned and characteristics of
  the task or method
• In the example
• Type of credit task-does-not-produce-value
• Localized to the find-next-slot task (a
  primitive task implemented by a LISP procedure)
• SIRRINE selects the fixed value production by
  task decomposition strategy.

16
Fixed Value Productionby Task Decomposition

• Existing task is divided into two methods a base
  method and an alternate method.
• The base method invokes a single task whose
  behavior is identical to the existing task.
• The alternate method invokes a new task
• Primitive that produces a single fixed value
• Example 430PM to 600PM on Tuesdays
• Applicability conditions on the alternate method
  require that it be invoked only in the same
  situation.
• Same defined by bindings in the trace that are
  referenced in the model of the existing task.

17
Model Zoom
18
Relevant Portions of the Model
19
Modification to the Model
20
ExampleWeb Browsing Agent

• A mock-up of web browsing software
• Based on Mosaic for X Windows, version 2.4
• Imitates not only behavior but also internal
  process and information of Mosaic 2.4

ps
???
html
pdf
txt
21
Example PDF Viewer

• The web agent is asked to browse the URL for a
  PDF file.
• Mosaic 2.4 not have any information about
  external viewers for PDF.
• The system cannot display the file.
• The user provides feedback indicating the command
  for the correct viewer.
• Adaptation Strategy Fixed Value Production by
  Primitive Modification

22
SIRRINE User Interface
23
REM Proactive Adaptation

• Reflective Evolutionary Mind
• Like SIRRINE, REM is a shell for adaptive
  software systems using TMK models.
• Unlike SIRRINE, REM is able to address new tasks.
• It can retrieve and adapt methods for known
  tasks, or it can build new methods from scratch.
• Off-the-shelf generative planning and
  reinforcement learning techniques are used to
  build new methods

24
REM Functional Architecture
Model
25
Task-Method-Knowledge Language (TMKL)

• A new, powerful formalism of TMK developed for
  REM.
• Uses LOOM, a popular off-the-shelf knowledge
  representation framework concepts, relations,
  etc.

REM models not only the tasks of the domain but
also itself in TMKL.
26
Sample TMKL Task

• (define-task communicate-with-www-server
• input (input-url)
• output (server-reply)
• makes
• (and
• (document-at-location (value server-reply)
  
• (value
  input-url))
• (document-at-location (value server-reply)
• 
  local-host))
• by-mmethod (communicate-with-server-method))

27
Sample TMKL Method

• (define-mmethod external-display
• provided (not (internal-display-tag (value
  server-tag)))
• series (select-display-command
• compile-display-command
• execute-display-command))

28
Decision Making in REM Q-Learning

• Popular, simple form of reinforcement learning.
• In each state, each possible decision is assigned
  an estimate of its potential value (Q).
• For each decision, preference is given to higher
  Q values.
• Each decision is reinforced, i.e., its Q value
  is altered based on the results of the actions.
• These results include actual success or failure
  and the Q values of next available decisions.

29
Q-Learning in REM

• Decisions are made for method selection and for
  selecting new transitions within a method.
• A decision state is a point in the reasoning
  (i.e., task, method) plus a set of all decisions
  which have been made in the past.
• Initial Q values are set to 0.
• Decides on option with highest Q value or
  randomly selects option with probabilities
  weighted by Q value (configurable).
• A decision receives positive reinforcement when
  it leads immediately (without any other
  decisions) to the success of the overall task.

30
ExampleDisassembly and Assembly

• Software agent for disassembly operating in the
  domain of cameras
• Information about cameras
• Information about relevant actions
• e.g., pulling, unscrewing, etc.
• Information about the disassembly process
• e.g., decide how to disconnect subsystems from
  each other and then decide how to disassemble
  those subsystems separately.
• Agent now needs to assemble a camera

31
Adaptation UsingRelation Mapping

• Requires a model for an existing agent which has
  a task similar to the desired task.
• e.g., disassembly is similar to assembly
• Effects (makes slot) of the two tasks must match
  except for one term, and that one term must be
  connected by a single relation.
• e.g., disassembly produces a disassembled state
• assembly produces an assembled state
• (inverse-of disassembled assembled) is known.
• Uses the relation to alter tasks and methods

32
Pieces of ADDAM which are key to Disassembly ?
Assembly
Disassemble
Plan Then Execute Disassembly
Adapt Disassembly Plan
Execute Plan
Hierarchical Plan Execution
Topology Based Plan Adaptation
Make Plan Hierarchy
Map Dependencies
Select Next Action
Execute Action
Select Dependency
Assert Dependency
Make Equivalent Plan Nodes Method
Make Equivalent Plan Node
Add Equivalent Plan Node
33
New Adapted Task inDisassembly ? Assembly
Assemble
COPIED Plan Then Execute Disassembly
COPIED Adapt Disassembly Plan
COPIED Execute Plan
COPIED Hierarchical Plan Execution
COPIED Topology Based Plan Adaptation
COPIED Make Plan Hierarchy
COPIED Map Dependencies
Select Next Action
INSERTED Inversion Task 2
Execute Action
COPIED Select Dependency
INVERTED Assert Dependency
COPIED Make Equivalent Plan Nodes Method
COPIED Add Equivalent Plan Node
INSERTED Inversion Task 1
COPIED Make Equivalent Plan Node
34
Task Assert Dependency

• Before
• define-task Assert-Dependency
• input target-before-node, target-after-node
• asserts (node-precedes (value
  target-before-node)
• (value target-after-node))
• After
• define-task Mapped-Assert-Dependency
• input target-before-node, target-after-node
• asserts (node-follows (value
  target-before-node)
• (value target-after-node)))

35
Task Make Equivalent Plan Node

• define-task make-equivalent-plan-node
• input base-plan-node, parent-plan-node,
  equivalent-topology-node
• output equivalent-plan-node
• makes (and
• (plan-node-parent (value
  equivalent-plan-node)
• 
  (value parent-plan-node))
• (plan-node-object (value
  equivalent-plan-node)
• 
  (value equivalent-topology-node))
• (implies (plan-action (value
  base-plan-node))
• (type-of-action
  (value equivalent-plan-node)
• 
  (type-of-action (value base-plan-node)))))
• by procedure ...

36
TaskInverted-Reversal-Task

• define-task inserted-reversal-task
• input equivalent-plan-node
• asserts (type-of-action
• (value equivalent-plan-node)
• (inverse-of
• (type-of-action
• (value
  equivalent-plan-node))))

37
Adaptation UsingGenerative Planning

• Does not require a pre-existing model
• Requires operators and a set of facts (initial
  state)
• Invokes Graphplan
• Operators Those primitive tasks known to the
  agent which can be translated into Graphplans
  operator language
• Facts Known assertions which involve relations
  referred to by the operators
• Goal Makes condition of main task
• Translates plan into more general method by
  turning specific objects into parameters
  propagating
• Stores method for later reuse

38
Adaptation UsingSituated Learning

• Does not require a pre-existing model
• Does not even require preconditions and
  postconditions of the operators
• Creates a method that
• Performs any action
• Checks whether the desired state is achieved
• If not, loops to the start.
• During execution, all decision making is done
  using Q-learning policy.
• Over time, the Q-learning mechanism selects
  actions that tend to lead to desirable results.

39
ADDAM Example Layered Roof
40
Roof Assembly
Situated Learning
Generative Planning
Relation Mapping
41
Modified Roof Assembly No Conflicting Goals
Situated Learning
Relation Mapping
Generative Planning
42
Combining Proactive Retrospective Adaptation in
REM

• Proactive adaptation techniques have been the
  primary focus of REM
• However, REM also has facilities for
  retrospective adaptation
• inherited from SIRRINE
• REM can use SIRRINE-style analysis of traces to
  localize an opportunity for adaptation to a
  particular subtask.
• It can then use a proactive technique to build a
  new method for that subtask.

43
Explanation

• As I have discussed, TMK models are useful for
  automated adaptation.
• This implies that they encode important knowledge
  about processes.
• This suggests that TMK may be an effective
  mechanism for explaining processes to human
  users.
• Some TMK research has investigated this idea.

44
Interactive KritikSelf-Explanation

• Objective Interactive explanation and
  justification for conceptual design of physical
  devices
• Input Functional specification for a device
• Output Model of a device that meets the
  specification and a graphical explanation of how
  the device was designed
• Knowledge Library of functional models of
  devices and a graphical model of the design
  process
• Reasoning Kritik2 (Goel, Bhatta, Stroulia,
  1997) performs case-based design. Interactive
  Kritik adds graphical presentation of the
  reasoning process and product.

45
Part of a behavior of an acid cooler
Water at 25º
Heated to 50º
By clicking on the transition, a user can jump to
the part of the acid cooling behavior that is
enabled by the heating of the water
46
That task is accomplished by a reasoning method.
The top level task of Interactive Kritik is
design.
The method decomposes the task into subtasks.
Some subtasks have methods that further decompose
them.
47
AHEADExplanation of Threats

• Analogical Hypothesis Elaborator for Activity
  Detection
• Objective Helping intelligence analysts
  understand and trust hypotheses about detected
  hostile activity
• e.g., organized crime, terrorism
• Input Hypothesis about hostile activity
  related evidence.
• Output Arguments for and/or against the
  hypothesis.
• Knowledge TMK models encode how hostile actions
  are performed and what they are intended to
  accomplish.
• Reasoning First, MAC/FAC (Forbus Gentner 1991)
  maps the hypothesis to a TMK model. Then, TMK
  simulation guides analysis of hypotheses using
  evidence.

48
AHEAD Functional Architecture
FIRE Analogy Server
Link Discovery Tools
Hypothesis/Model Mapping
TMK Models (Qualitative, Functional)
Trace Extractor
Evidence Extraction Tools, Existing
Knowledge-Bases
Model Trace
TIA, Assorted EELD Tools, etc.
Argument Generator
GUI
49
AHEAD User Interface
Statement of the hypothesis (input)
Red and black icons indicate qualitative
certainty for arguments and evidence
Arguments against the hypothesis have missing or
contradictory evidence
Arguments for the hypothesis are backed by
evidence
Hyperlinks to original sources for evidence.
A key allows users to quickly see what each icon
means.
50
Preliminary User Study

• Partial implementation handmade output files.
• Tested the interface and content of AHEAD.
• In some trials, users were given hypothesis
  evidence
• i.e., inputs to AHEAD
• In other trials, users also given arguments for
  and arguments against.
• Users with arguments showed better performance.
• Difference in error in judgment statistically
  significant.

Metric WithArgument WithoutArgument
Elapsed Time 518 555
Confidence 7.40 6.86
Error in judgment 1.83 3.01
Error in confidence 1.70 2.26
51
Summary

• Retrospective adaptation SIRRINE
• Encounters a constraint during execution
• Uses trace to find a candidate for modification
• Uses model to perform modification
• Experiments in meeting scheduling and web
  browsing (etc.)
• Proactive adaptation REM
• User requests a new task
• REM retrieves similar task and adapts the method
  for that task
• Or REM uses Graphplan or Q-learning to build a
  new method
• Can also use these techniques retrospectively on
  subtasks
• Self-explanation Interactive Kritik
• Presentation of design process and product using
  models
• Explanation of threats AHEAD
• Models ? traces ? arguments for and arguments
  against hypotheses.
• Arguments help analysts understand and trust
  hypotheses.

52
Backup Slides
53
SIRRINE
54
Model Knowledge
55
Model Tasks and Methods
56
Model
57
Tasks and Methodsof Web Agent
Process URL
Process URL Method
Communicate with WWW Server
Display File
Communicate with WWW Server Method
Display File Method
Request from Server
Receive from Server
Interpret Reply
Display Interpreted File
External Display
Internal Display
Execute Internal Display
Select Display Command
Compile Display Command
Execute Display Command
58
Web Agent Adaptation
...
External Display
Select Display Command
Compile Display Command
Execute Display Command
...
External Display
Compile Display Command
Execute Display Command
Select Display Command
Select Display Command Base Method
Select Display Command Alternate Method
Select Display Command Base Task
Select Display Command Alternate Task
59
REM/TMKL
60
Tasks in TMKL

• All tasks can have input output parameter lists
  and given makes conditions.
• A non-primitive task must have one or more
  methods which accomplishes it.
• A primitive task must include one or more of the
  following source code, a logical assertion, a
  specified output value.
• Unimplemented tasks have neither of these.

61
TMKL Task

• (define-task communicate-with-www-server
• input (input-url)
• output (server-reply)
• makes
• (and
• (document-at-location (value server-reply)
  
• (value
  input-url))
• (document-at-location (value server-reply)
• 
  local-host))
• by-mmethod (communicate-with-server-method))

62
Methods in TMKL

• Methods have provided and additional result
  conditions which specify incidental requirements
  and results.
• In addition, a method specifies a start
  transition for its processing control.
• Each transition specifies requirements for using
  it and a new state that it goes to.
• Each state has a task and a set of outgoing
  transitions.

63
Simple TMKL Method

• (define-mmethod external-display
• provided (not (internal-display-tag (value
  server-tag)))
• series (select-display-command
• compile-display-command
• execute-display-command))

64
Complex TMKL Method

• (define-mmethod make-plan-node-children-mmethod
• series (select-child-plan-node
• make-subplan-hierarchy
• add-plan-mappings
• set-plan-node-children))
• (tell (transitiongtlinks make-plan-node-children-mm
  ethod-t3
• equivalent-plan-nodes
• child-equivalent-plan-nod
  es)
• (transitiongtnext make-plan-node-children-mm
  ethod-t5
• make-plan-node-children-mm
  ethod-s1)
• (create make-plan-node-children-terminate
  transition)
• (reasoning-stategttransition
  make-plan-node-children-mmethod-s1
• 
  make-plan-node-children-terminate)
• (about make-plan-node-children-terminate
• (transitiongtprovided
• '(terminal-addam-value (value
  child-plan-node)))))

65
Knowledge in TMKL

• Foundation LOOM
• Concepts, instances, relations
• Concepts and relations are instances and can have
  facts about them.

Knowledge representation in TMKL involves LOOM
some TMKL specific reflective concepts and
relations.
66
Some TMKLKnowledge Modeling

• (defconcept location)
• (defconcept computer
• is-primitive location)
• (defconcept url
• is-primitive location
• roles (text))
• (defrelation text
• range string
• characteristics single-valued)
• (defrelation document-at-location
• domain reply
• range location)
• (tell (external-state-relation
• document-at-location))

67
Sample Meta-Knowledge in TMKL

• relation characteristics
• single-valued/multiple-valued
• symmetric, commutative

• relations over relations
• external/internal
• state/definitional

• generic relations
• same-as
• instance-of
• inverse-of

• concepts involving concepts
• thing
• meta-concept
• concept

68
Planning and REM

• REM is not a planning system, its an agent
  environment.
• Agents in REM dont just decide what to do, they
  actually do things.
• They may base their next action on the results of
  past actions.
• Once they have taken an action, they may not be
  able to undo that action.
• REM consists of two major modules execution of
  agents and adaptation of agents.
• However, research on REM has involved many
  planning issues.
• Agents in REM are represented as hierarchies of
  tasks and methods thus execution of agents
  resembles HTN planning.
• One mechanism for adaptation of agents is
  traditional generative planning.
• Agents encoded in REM may perform planning.
• An agent in REM can act in a deterministic
  logically-defined simulated environment such
  agents blur the distinction between deciding what
  to do and doing things.

69
REM vs.Derivational Analogy

• REM adapts models of tasks and methods.
• Derivational analogy generally assumes some sort
  of universal process (e.g., generative planning)
  and only needs to represent and reason about key
  decision points.
• Advantage of derivational analogy Models not
  needed traces alone enable reuse.
• Advantage of REM Applicable to problems for
  which a universal process is not appropriate
  (e.g., 6 board roof example takes days using
  planning Q-learning).
• REM demands more knowledge but makes effective
  use of that additional knowledge.

70
REM andCase-Based Reasoning (CBR)

• Given a task, REM retrieves a method, applies it,
  and then (if necessary) adapts it. This process
  is a form of CBR.
• Most CBR projects, however, adapt solutions not
  processes. Some problems require the latter.
• Adaptation of processes can enable extending the
  efficiency benefits of CBR to problems which the
  case library does not directly address.

71
Recursion and Looping
72
REM vs.Case-Based Adaptation

• REM reasons about and adapts an entire reasoning
  process.
• Case-based adaptation restricts adaptation to one
  portion of a case-based process adaptation.
• Being more focused is a substantial advantage for
  case-based adaptation.
• However, for problems which require adaptation of
  different sorts of reasoning processes, it is
  useful to have models of these processes, as in
  REM.

73
Q-Learning in REM

• Decisions are made for method selection and for
  selecting new transitions within a method.
• A decision state is a point in the reasoning
  (i.e., task, method) plus a set of all decisions
  which have been made in the past.
• Initial Q values are set to 0.
• Decides on option with highest Q value or
  randomly selects option with probabilities
  weighted by Q value (configurable).
• A decision receives positive reinforcement when
  it leads immediately (without any other
  decisions) to the success of the overall task.

74
Monkey Bananas Tower of Hanoi Hybrid
75
AHEAD
76
Key Ideas for AHEAD
Objective Helping analysts understand and trust
hypotheses about detected hostile activity
Knowledge Functional models encode how hostile
actions are performed and what they are intended
to accomplish. Reasoning Analogy maps
hypotheses to models. Functional models guide
analysis of hypotheses in the context of
evidence. Product Structured arguments that
support and/or refute the given hypothesis.
77
ReasoningFIRE Analogy Server

• Unified system for performing analogical
  reasoning from Ken Forbus group at Northwestern
  University.
• SME Structure-Mapping Engine
• Produces a mapping between source and target
  cases.
• Evaluates similarity based on these mappings.
• Recommends candidate inferences, i.e., new
  elements for the target that would be analogous
  to elements in the source.
• MAC/FAC Many are Called / Few are Chosen
• Searches a library of cases for those that match
  a given target.
• MAC step uses fast superficial matching to
  identify many candidates.
• FAC step uses SME for more precise matching.
• Additional related tools, e.g., a GUI knowledge
  acquisition tool based on SME.

78
Illustrative ExampleExtracted Trace

• The trace shows
• Steps in the model that were taken
• Values the parameters were bound to
• Pieces of evidence that support the steps and
  bindings.

TMK Model
Evidence
Architecture
79
AHEAD Functional Architecture Structured Argument
A hypothesis is a potential conclusion about what
happened in the world.
An argument against asserts that some part of the
relevant model was not performed.
An argument for is a asserts that some part of
the relevant model was performed.
Hypothesis
...
...
Argument For
Argument Against
Argument Against
Argument For
Other argument against simply indicate a lack of
evidence.
Evidence
Evidence
Evidence
Some arguments against have explicit links to
evidence that supports them.
All arguments for are supported by references to
the original evidence.
Dashed lines in this figure indicate a connection
plus a qualitative confidence level.
80
Other
81
Comparison
Autognostic SIRRINE REM
Why To repair agents which fail to address the task for which they were designed To add new capabilities to an agent. To add new capabilities to an agent.
What Modifications to the model Modifications to the model and addition and modification of primitive computations Modifications to the model and addition and modification of primitive computations
How Collection of strategies which make specific types of changes. Collection of strategies which make specific types of changes. Collection of strategies which make specific types of changes, plus generative planning and reinforcement learning.
When After execution After execution Before and after execution
82
Bibliography

• Chandrasekaran, B. (1986). Generic tasks in
  knowledge-based reasoning High-level building
  blocks for expert systems design. IEEE Expert
  13, 23-30.
• Gentner, D. Forbus, K. (1991). MAC/FAC A model
  of similarity-based retrieval. Proceedings of the
  Thirteenth Annual Conference of the Cognitive
  Science Society (pp. 504-509). Chicago, IL
  Lawrence Earlbaum.
• Goel, A.K., Bhatta, S., Stroulia, E. (1997).
  Kritik An Early Case-Based Design System. In
  Maher, M. and Pu, P. (Eds.) Issues and
  Applications of Case-Based Reasoning in Design
  (pp. 87-123), Mahwah, NJ Erlbaum.
• Goel, A.K. Murdock, J.W. (1996). Meta-Cases
  Explaining Case-Based Reasoning. Proceedings of
  the Third European Workshop on Case-Based
  Reasoning. Lausanne, Switzerland Springer.
• Griffith, T.W. (1999). A computational theory of
  generative modeling in scientific reasoning. PhD
  Thesis. College of Computing, Georgia Institute
  of Technology.
• Liver, B. Allemang, D. (1995). A functional
  representation for software design. International
  Journal of Software Engineering and Knowledge
  Engineering, 52, 227-269.
• Murdock, J.W. (2001). Self-Improvement through
  Self-Understanding Model-Based Reflection for
  Agent Adaptation. PhD Thesis. College of
  Computing, Georgia Institute of Technology.
• Murdock, J.W., Aha, D.W., Breslow, L.A. (2003).
  Case-based argumentation via process models. To
  appear in Proceedings of the Fifteenth
  International Conference of the Florida
  Artificial Intelligence Research Society. St.
  Augustine, FL AAAI Press.
• Murdock, J.W. Goel, A.K. (2001). Learning about
  Constraints by Reflection. Proceedings of the
  Fourteenth Canadian Conference on Artificial
  Intelligence (AI'01), Ottawa, Ontario
• Sembugamoorthy, V. Chandrasekaran, B. (1986).
  Functional representation of devices and
  compilation of diagnostic problem-solving
  systems. In J. Kolodner C. Reisbeck (Eds.)
  Experience, Memory, and Learning (pp. 47-73).
  Mahwah, NJ Erlbaum 1986.
• Stroulia, E., Goel, A.K. (1995). Functional
  representation and reasoning in reflective
  systems. Applied Intelligence, 9, 101-124.