Task Learning in COLLAGEN

1
Task Learning in COLLAGEN

• The COLLAGEN Architecture Task Learning from
  Demonstrations
• Work _at_ Mitshubishi Electric Research Labs
• Andrew Garland, Neal Lesh, Kathy Ryall, Charles
  Rich,
• Candy Sidner
• Carnegie Mellon University, 2001

2
Outline

• The COLLAGEN Architecture
• P1 COLLAGEN Applying Collaborative Discourse
  Theory to Human-Computer Interaction
• Learning Task Models
• P2 Learning Task Models from Collaborative
  Discourse
• Add refinement regression testing
• P3 Learning Hierarchical Task Models by Defining
  and Refining Examples
• Adding guessers
• P4 Interactively Defining Examples to be
  Generalized
• Discussion pros, cons, questions

3
COLLAGEN

• COLLAGEN COLLaborative AGENt
• Based on SharedPlans discourse theory (Grosz
  Sidner)
• Not the classical dialog-system view agent
  human collaborate, and they both interact with
  the application
• 4 agents presented VCR, SymbolEditor, GasTurbine
  agent, home thermostat (kind-of toy domains)

4
COLLAGEN (contd)

• Dialog Management architecture
• Discourse state
• Focus stack (stack of goals)
• Plan tree for each of them
• Actions primitive / non-primitive
• Recipes specification of goal decompositions
• Partially ordered steps, parameters, constraints,
  pre- and post-conditions
• Updating the discourse state 5 cond
• Plan recognition

5
Learning Task Models from Collaborative Discourse
2

• Starting Point more difficult for people to
  deal with abstractions in the task than to
  generate and discuss examples
• Programming-by-Demonstration approach
• Infer task models from partially-annotated
  examples of task behavior.
• Similarities with Helpdesk Call Center
• CallCenter idea learn from watching traffic
• Richly annotate traffic / recent EARS stuff
• Learn task structure from annotated traffic

6
Learning Task Models (contd)

• Annotation Language
• e, S, optional, unordered, unequal
• Q how powerful is this task representation ?
• fully annotating would be burdensome
• Learning alignment, optionals, orders
  propagators
• BIAS for learning
• Alignment Disjoint step assumption
• Alignment Step type assumption.
• Q Hmm, not sure I got this
• Propagators Suggested parameter preference bias
  ( occams razor)

7
Learning Task Models Experiments.

• How
• Start from 2 task models
• Generate examples, randomize
• Relearn models, see what you get
• Results
• Optional did not get much action it figures,
  its probably the easiest to learn
• Equality seems to buy a lot and this is good !
• Learning is strongly influenced by the order of
  examples
• Discussion
• Not adequate for direct use
• Mention of the online flavor

8
Learning HTN by defining and refining examples

• Created a development environment which
  integrates the learning techniques with
• Defining Refining examples
• Regression testing (needed if manual edits are
  allowed)
• They esentially give a management process for the
  development of task models fig. 3
• Q Is there any reason for Starting Set of
  Actions ?
• Q The whole things looks really like a
  storyboard, but is there anything really new here
  ?

9
Interactively Defining Examples to be Generalized

• NEW Guessers
• Guessers suggest to the user what annotations
  might be helpful
• Organized in committees to improve robustness
• Knowledge sources
• Other examples
• Current generalization
• The inference techniques active learning
• Raw data
• Domain Theory
• Heuristics

10
So what do you think ?

• Is it worth it ? When ?
• Does the conjecture hold ?
• How about when you collect the examples ? (ala
  CallCenter)